scieee AI-readable full text Open interactive document viewer

Patient specific models of clotting

Ripoll Oliveras, Albert

Abstract

There is an increasing impact of cardiovascular diseases on the population worldwide that have lead them to be the first cause of global mortality. Most of these cardiovascular diseases evolve to high fatality rates when undesired clots are generated. Blood clotting is a complex phenomenon influenced by many biophysical mechanisms and factors happening simultaneously. All these interconnected mechanisms produce a huge interpatient variability. However, patient treatments remain largely the same. This work is in the scope of coronary heart disease precision medicine and it consists on a study of differential clotting patterns observed in various patient specific coronary anatomies in order to identify distinct biophysical mechanisms driving these patterns. The overarching objective of the thesis is that if distinct clotting behaviors can be dissected and identified, treatments can be tailored and targeted with more specificity. In order to identify and validate clotting behaviors, a wide range of techniques were required including biological experimentation, benchtop model development and prototyping, computational fluid dynamic simulation as well as different machine learning techniques for cluster matching responses between computational and experimental models. With these tools, hypotheses previously developed from computational fluid dynamics simulations were probed through some ex-vivo experiments. From irregular blood vessel geometries, a map of the clotting space was developed, what lead to the visualization of several clusters with similar behaviors and critically, a novel cluster of atherosclerotic geometries that are particularly prone to severe clotting, yet which are not currently captured clinically. Moreover, the different clusters were predicted from clinical variables avoiding the expensive computations of the fluid dynamic simulations. This fact gave rise to the found that new clinical variables which provide valuable insights about blood flow need to be created and validated in order to achieve precise diagnoses considering hemodynamics.

Full text

Patient specific models of clotting 1 Treball Final del Grau en Enginyeria Física Patient specific models of clotting Albert Ripoll Oliveras supervised by Prof. Kumaran Kolandaivelu co-directed by Prof. Blas Echebarria Boston, May 2018. Patient specific models of clotting 2 Abstract There is an increasing impact of cardiovascular diseases on the population worldwide that have lead them to be the first cause of global mortality. Most of these cardiovascular diseases evolve to high fatality rates when undesired clots are generated. Blood clotting is a complex phenomenon influenced by many biophysical mechanisms and factors happening simultaneously. All these interconnected mechanisms produce a huge interpatient variability. However, patient treatments remain largely the same. This work is in the scope of coronary heart disease precision medicine and it consists on a study of differential clotting patterns observed in various patient specific coronary anatomies in order to identify distinct biophysical mechanisms driving these patterns. The overarching objective of the thesis is that if distinct clotting behaviors can be dissected and identified, treatments can be tailored and targeted with more specificity. In order to identify and validate clotting behaviors, a wide range of techniques were required including biological experimentation, benchtop model development and prototyping, computational fluid dynamic simulation as well as different machine learning techniques for cluster matching responses between computational and experimental models. With these tools, hypotheses previously developed from computational fluid dynamics simulations were probed through some ex-vivo experiments. From irregular blood vessel geometries, a map of the clotting space was developed, what lead to the visualization of several clusters with similar behaviors and critically, a novel cluster of atherosclerotic geometries that are particularly prone to severe clotting, yet which are not currently captured clinically. Moreover, the different clusters were predicted from clinical variables avoiding the expensive computations of the fluid dynamic simulations. This fact gave rise to the found that new clinical variables which provide valuable insights about blood flow need to be created and validated in order to achieve precise diagnoses considering hemodynamics. Patient specific models of clotting 3 Acknowledgements First and foremost, I would like to thank Dr. Kumaran Kolandaivelu for giving me the opportunity to work side by side with him at the Clinical Research Center (CRC) during these months. Thanks to his advice and mentoring, this stage has been a constant learning. It has been a pleasure meeting him both, professionally and personally. Next, I would like to thank Catherine Ricciardi and O’Mara Taylor for taking care of CRC and making it a very comfortable place to work at. Gràcies a tot l’equip del CRC que m’ha acompanyat aquests mesos: Marc, Jose, Patricia, Bruno, Laura, Anna, Martí, Ferran i Adrià. Tots i cadascun d’ells, a la seva manera, ha aportat quelcom a aquest treball. Tot i la intensitat de treball i els horaris, a vegades exigents, han fet de l’estada al CRC una època inoblidable. Agrair també a la Maria la seva incansable ajuda a tots els experiments i la demostració de perseverança malgrat els diversos resultats adversos que es van obtenir al llarg de l’experimentació. También quiero agradecer el profesor Blas Echebarria, quien ha tutelado la tesis desde la distancia, su ayuda y apoyo a lo largo del proyecto y a la hora de redactar esta memoria. Vull agrair, igualment, el CFIS i el seu programa de mobilitat conjuntament amb la Fundació Cellex que el patrocina per tot allò que han aportat perquè es fes possible aquest treball fi de grau. Finalment, però no menys important, moltes gràcies a la família. Tot i estar a milers de quilòmetres lluny, han estat aquí, al meu costat, recolzant-me i oferint-me consell sempre que ho necessitava. Especialment als meus pares, pel seu suport i amor incondicional. Sense ells, tot això no hauria estat impossible. Patient specific models of clotting 4 Table of contents ABSTRACT _________________________________________________ 2 ACKNOWLEDGEMENTS ______________________________________ 3 TABLE OF CONTENTS _______________________________________ 4 1. INTRODUCTION _________________________________________ 6 1.1. Project framework ...................................................................................... 6 1.2. Motivation and introduction to Cardiovascular Diseases ........................... 6 1.3. Previous work ............................................................................................ 9 1.4. Project overview ....................................................................................... 11 2. THEORETICAL FOUNDATIONS ___________________________ 13 2.1. Atherosclerosis ........................................................................................ 13 2.2. Introduction to clotting and mechanisms of hemostasis .......................... 14 2.2.1. Platelet plug formation ................................................................................. 14 2.2.2. Coagulation cascade ................................................................................... 15 2.3. Hemodynamics as clotting driver ............................................................. 16 3. METRICS DEFINITIONS _________________________________ 18 3.1. Geometries .............................................................................................. 18 3.2. Flow parameters ...................................................................................... 19 3.2.1. Recirculation (Ujk) ........................................................................................ 20 3.2.2. Mass transfer: Peclet Number (Pe) ............................................................. 21 3.2.3. Tangential Velocity (Utn) .............................................................................. 22 3.2.4. Wall Shear Stress (WSS) ............................................................................ 22 3.2.5. Vorticity (Vk) ................................................................................................ 23 3.3. Clinical variables ...................................................................................... 24 4. EXPERIMENTAL WORK _________________________________ 26 4.1. Overview .................................................................................................. 26 4.2. Methodology: Flow loop experimentation ................................................ 26 4.2.1. Introduction to flow loop model .................................................................... 26 4.2.2. Surface treatments and coatings ................................................................. 28 4.2.3. Blood treatments ......................................................................................... 29 4.2.4. Clotting test ................................................................................................. 29 4.2.5. Kuminator’s calibration ................................................................................ 30 4.2.6. Results analysis. Lactate Dehydrogenase (LDH) and Hemoglobin assays . 31 4.3. Initial experiments .................................................................................... 32 4.3.1. Geometry selection ..................................................................................... 32 4.3.2. Results ........................................................................................................ 32 Patient specific models of clotting 5 4.4. Specific geometries test ........................................................................... 33 4.4.1. Geometry selection ..................................................................................... 33 4.4.2. Results ........................................................................................................ 36 4.5. Mapping clotting space and classification of stenotic geometries ............ 41 4.6. Attempts to recover flow loop signal ........................................................ 44 4.7. Delimitation of clotting clusters ................................................................. 45 4.7.1. Geometry selection ..................................................................................... 46 4.7.2. Results ........................................................................................................ 48 5. COMPUTATIONAL DATA ANALYSIS _______________________ 49 5.1. Overview .................................................................................................. 49 5.2. Predicting models of flow parameters ...................................................... 49 5.3. Prediction of clotting clusters ................................................................... 51 6. CONCLUSIONS ________________________________________ 54 7. REFERENCES _________________________________________ 55 Patient specific models of clotting 6 1. Introduction 1.1. Project framework This project has been developed at the Clinical Research Center (CRC), a department housed in the Institute for Medical Engineering and Science (IMES), at Massachusetts Institute of Technology’s (MIT). The CRC aims to ease the translation of new scientific discoveries into practical clinical applications to advance human health. Researchers at CRC are provided with a wide variety of resources to help them in their research for medical innovation and execute studies safely and effectively, from early study planning and clinical consultation to specialized laboratory spaces. In the past few years, CRC projects have been focusing on precision medicine. This concept refers to all those disease treatment that take into account the variability between individuals, such as their genetics, environment or lifestyle. This implies that the optimal required treatment might differ between two individuals affected by the same disease. Therefore, precision medicine is a medical model that proposes the customization of healthcare, with medical decisions, treatments, practices or products being tailored to the individual patient. The fact of tailoring medical treatment to the individual characteristics of each patient does not literally mean the creation of new drugs or medical devices that are unique to a single patient, but rather the ability to classify individuals into subpopulations that differ in their susceptibility to a particular disease or in their response to a specific treatment. This concept is one of the central pillars of this thesis as the research has been rooted in the classification and subgrouping of different clotting patterns. The thesis has been carried out under the supervision and mentoring of Dr. Kumaran Kolandaivelu, Medical Director of the CRC. Dr. Kolandaivelu received his Bachelor of Science (BS) and Master of Science (MS) in Mechanical Engineering, and a PhD in Medical Engineering & Medical Physics at MIT. Following a Medical Doctorate Program from the Harvard-MIT Division of Health Sciences and Technology (HST), and he completed his training in internal medicine and cardiology at Brigham and Women’s Hospital (BWH), where he is now a board-certified physician in the BWH Cardiovascular Division. 1.2. Motivation and introduction to Cardiovascular Diseases The field in which the CRC invests most of its resources is the study of cardiology and blood coagulation. This interest rises from the fact that heart diseases and strokes are the main causes of death worldwide, as seen in Figure 1. Although in most of the developed countries Cancer has surpassed Ischaemic Heart diseases (IHD) as the leading cause of death, globally it is still the main cause. Patient specific models of clotting 7 Figure 1. Ischaemic heart diseases and strokes are the main global causes of death, and both have cardiovascular origins (Source: World Health Organization). More specifically, the diverse projects recently developed -comprising this oneat the CRC focus on the coagulation process in the coronary arteries. The coronary arteries are the arteries that transport blood into and out of the cardiac muscle. These arteries originate at the base of the aorta and end in a ramified net of tributaries that extend throughout the myocardial tissue. The malfunction of these arteries may lead to several coronary artery disease (CAD) such as angina and myocardial infarction (MI) -commonly known as heart attack-, ischaemic stroke, heart failure, valvular heart disease, heart arrhythmia including sudden cardiac death, and congenital heart defects among many others diseases [1]. Many of which are related to a disease called atherosclerosis. The process of atherosclerosis refers to certain conditions in which plaque is generated inside the blood vessels. Plaque or atheroma is an accumulation of fat, cholesterol, calcium and some other substances in and on the artery walls. Advanced forms of atherosclerosis usually result in a reduction of the artery’s inner cross section, known as stenosis, creating a geometrical perturbation or bump. When plaque deposited in the heart coronary arterial wall becomes unstable and ruptures, a phenomenon known as thrombosis starts, concluding in a blood clot that causes either a reduction or an acute occlusion of blood flow (Figure 2). Consequently, blood rich in oxygen and nutrients is prevented from reaching the cardiac tissue and supplying it (ischaemia) therefore all muscle tissue below the blockage can die. This event may lead to most of the CAD mentioned before. Patient specific models of clotting 8 Figure 2. Coronary artery atherosclerosis evolution (Source: American Heart Association, [2]). The formation of a blood clot is the result of a complex process in which the blood changes from liquid to gel with the objective of stopping the loss of blood when a blood vessel is injured, the cessation of the bleeding is called hemostasis. The mechanism of coagulation involves several agents, from different molecules and ions to platelets and red blood cells. Coagulation is vital for life as long as it works correctly, however, if clots appear in undesired locations they can lead to fatal results. The formation of a blood clot inside a blood vessel, obstructing the flow through the circulatory system is known as thrombosis. In 1856, Virchow described a triad of factors that are thought to contribute to thrombosis (Figure 3). The first component of the triad, hypercoagulability state, involves alterations in the constitution of the blood and therefore an increase in thrombosis risk. Variations in the concentration of molecules or cells may produce for example hyperviscosity. The second category, injuries and/or trauma to endothelium -inner layer of the vessel wall-, may trigger coagulation when specific molecules located on the subendothelium layer, such as collagen and tissue factor, are exposed to blood. Finally, the last category is related to the alterations in normal blood flow and the velocity field of the blood inside the vessels called hemodynamics. For example, if laminar flow is interrupted, stasis brings platelets in contact with the endothelium allowing coagulation factors to accumulate and enhancing clotting. Figure 3. Virchow’s triad representation. Patient specific models of clotting 9 Therefore, thrombosis can be defined as a biophysical multifactorial process that can be modeled by taking into account these three core conditions. If two of the conditions are frozen, the third one can be studied and deeply understood. Being aware of the danger and fatality of CAD, the ultimate goal of the present thesis is to help diminish the frequency and mortality rate of myocardial infarctions through gaining insight into biophysical mechanisms that drive blood clotting in the atherosclerotic geometries. 1.3. Previous work This project is part of a research line that aims to study the interactions between the biological and the fluid dynamic process that occur in the coronary arteries and can lead to thrombosis with catastrophic consequences. This research line was born when consistent and opposites clotting pattern were observed in two different idealized coronary artery geometries. Different geometries extracted from clinical Optical Coherence Tomography (OCT) imaging were analyzed and used to create a generalized model. It was found that most of them had a similar bean-shaped cross section, therefore, an idealized mathematic curve of the cross section was generated and used in the simulations (Figure 4). Figure 4. Left: Patient OCT of an atherosclerotic artery (Source CRC data). Right: Idealized mathematic cross section curve. In the 3D extrapolation (Figure 5), two bump length were chosen as they represented two physiologically relevant geometries maintaining, although, the same typical cross section. The lengths were 6.0 and 1.5 mm. The long bump represents a generalization of lumen geometries typical from stenosis, while the second, the short one, is a usual geometry of a stent strut or spot of calcium. Patient specific models of clotting 16 Then, in both pathways, several coagulation factors, present in blood, begin to chemically react (see Figure 10) resulting in the formation of prothrombin activators. The prothrombin activators catalyzes the conversion of prothrombin to thrombin. Finally, thrombin turns fibrinogen into fibrin fibers that create a mesh stabilizing and aggregating platelets, blood cells and other substances forming a clot and concluding the secondary hemostasis [11]. It is important to remark that after blood vessel damaging, both clotting pathways begin at same time and that also a reciprocal influence exists between coagulation cascade and platelets activation. Therefore, usually, they are all taking place almost simultaneously but these mechanisms differ in their natures and their timings. However, if pathological, one specific mechanism is the leading cause of clotting. 2.3. Hemodynamics as clotting driver Hemodynamics refers to the dynamics of blood flow and explains the physical laws that govern the flow of blood in the blood vessels. According to Virchow’s triad, hemodynamic variations can contribute to thrombosis. In other words, blood flow alterations can potentially lead to coagulation inside the blood vessels. Blood is constantly flowing through vessels under different conditions, generating several physical events. Fluid dynamical phenomenon like vorticity, shear stress, helicity, the infinity of possible velocity profiles or internal forces caused by blood particles can interact with biological processes in multiple ways. In some of them, as clotting driver. Blood flow is the responsible for reactive substances transportation. Flow can accumulate them into certain regions, where the vessel might be damaged, but also it can remove and wash them away from the critical point by flow convection. In this case, if flow convection is stronger than the diffusion of the released tissue factor or other subendothelium substances, it can prevent the clot formation. Therefore, flow can be both, procoagulant and anticoagulant, depending on specific biology mechanisms [12]. Another clear example of hemodynamics influence in coagulation is the flow-dependence in the formation of platelets plug. In a similar way, flow affects to platelets adhesion because platelets receptors need to attach to vWF and collagen in the vessel wound. Moreover, platelets activation and vWF interaction depend closely on blood velocity profile, especially on shear stress [13]. Platelets can be activated at some vessel point by shear stress and later on be confined in a reduced region generating there and undesired clot as all the clotting factors are also confined in the same region due to flow characteristics [14]. Patient specific models of clotting 17 Figure 11. Biophysical interaction in an artery obstructed by a stenosis bump. Left: Biological mechanisms taking place are represented. Right: Blood flow trajectory lines distorted by the bump. (Source: M. Tomàs). Figure 11 is a representation of the interaction between the biologic and the physic worlds. An accumulation of different blood cells and agents that is highly probable to trigger coagulation mechanisms is found at a biological level. However, at the physical level, distortions in the trajectory lines of blood because of the obstacle can be observed. Coagulation is a biophysical process that depends on the behavior and interaction of the two worlds. Thanks to computational simulations flow characteristics can be known and can lead to the emergence of biophysical hypotheses. Patient specific models of clotting 18 3. Metrics definitions This thesis revolves around three main metrics: the geometric factors, the flow parameters and the clinical variables. In this section, the different metrics are accurately defined. Throughout the whole project, the interconnection of these metrics and also the relation with the biological results have been on the spotlight. 3.1. Geometries As it has been discussed through the last chapters, geometric irregularities in blood vessels may cause variability in the clotting patterns. Therefore, a mathematic model able to define geometric irregularities was needed. In such a way that all the artery geometries were defined by the same geometrical metrics that contained great precision and embraced almost all the possibilities. This model was stablished by previous researchers in order to be able to generate the coronary artery simulations. The geometries consist of a coronary artery of 3 mm diameter with a geometric irregularity, often called bump, which creates a narrowed or stenotic cross section. Each stenotic cross section is described by three parameters (see Figure 12):  e: Defines the distance between the lowest point of the circular cross section and the lowest point of the stenotic profile. Similar to an offset. The smooth transition to the 3 mm is guided by a hyperbolic function.  h: Defines the height of the bump without considering e. The curve is aided by a logit-normal curve.  w: It is the ratio between the area of the core of the bump divided by the height h. Figure 12. Three geometric parameters define a stenotic cross section (Source: A. Romeu). On the other hand, the bump has also two geometric parameters that describes the irregularity in the axial direction. The cross section profile adapts to the circular diameter through a Bezier curve of third order with a characteristic length defined by the parameters (see Figure 13): Patient specific models of clotting 19  LA: Defines upstream length.  LB: Defines downstream length. Figure 13. Two extra geometric parameters together with the cross section parameters totally define the longitudinal section. Finally an extra non-geometrical parameter is added to this set of metrics, the flow -Q-, in order to fully define all the parameters needed for the computational simulations. Combinations of these five parameters are selected to generate the different geometries and explore the clotting space. Its values are realistic descriptions of authentic lumen geometries. Table 1. Possible values of each geometric feature. A Sobol quasi-random sequence was used to choose feature combinations once the physically impossible or more improbable cases were discarded. The result was a total of 230 geometries simulated. From now on, each geometry will be referred as a number ranging from 1 to 230 that corresponds to the position in which they appear in the general compilation but 1.5 and 6.0 mm bumps. The table with all the simulations and geometries can be consulted in the Appendix, an example can be seen in Table 2. Table 2. Geometric parameters of the two initial bumps. 3.2. Flow parameters In the introduction, it was explained that hemodynamic changes are responsible of the variability in clotting patterns. However, it is not known which flow features have direct consequences on the patterns nor the way they interact with coagulation factors. Some hypotheses arose from the comparison between the velocity fields of the two key experiments -6.0 and 1.5 mm bumps- and five different physical flow factors were thought Patient specific models of clotting 20 to affect the polarized patterns. One flow parameter is thought to hence front clots while the other four seem to be closely related to side clots. A quantifiable metric for each flow parameter was computed for all geometries by a former researcher through a software of computational fluid dynamic simulator (STAR CCM+) followed by, in some occasions, a mathematical analysis. The quantitative metrics are crucial in order to be able to compare hundreds of simulations without the expensive requirement of having a plot for every single one. All metrics values that appear throughout this work are in International System of Units. Two thresholds, using 6.0 and 1.5 mm bumps flow metrics, were stablished to compare and evaluate simulations results. If a flow metric overpasses a threshold, such clotting pattern is expected to appear. The reader can go to Figure 7 to see the thresholds’ portrayal. 3.2.1. Recirculation (Ujk) The first hypothesis is related to recirculation. As seen in Figure 14, recirculation appears only in the short bump, therefore it seems to be responsible of the coagulation in the downstream side of the bump. It is needed to note that in the 1.5 mm bump, since the flow used experimentally spins back and forth in order to obtain a pulsatile flow, clots are formed not only downstream but also upstream. In a unidirectional flow, as in coronary arteries, the clot would only appear where recirculation takes place, downstream. This phenomenon commonly appears in geometric irregularities that causes a sudden expansion of the flow leading to a velocity deceleration and creating an adverse pressure gradient. If this negative pressure is high enough, the velocity deceleration can eventually create a backward flow. Therefore, recirculation generates a slow flow region where the products and factors that lead to clotting are not flushed out downstream and can concentrate, helping the process of coagulation. Figure 14. Velocity field in the longitudinal plane -“j-k” planefor 1.5 and 6.0 mm bumps. In 1.5 mm bump recirculation can be observed in the downstream side of the bump. The 6.0 mm bump has a smooth transition of shape while the 1.5 geometry changes sharply generating flow separation. Patient specific models of clotting 21 3.2.2. Mass transfer: Peclet Number (Pe) A common practice in the study of fluid mechanics is the use of experimental data in a dimensionless form as it helps to reduce the number of variables and maintain all the information. The classical example is the Reynolds number based on the ratio of inertial forces to viscous forces within a fluid. This dimensionless number helps to predict if fluid flow situations are dominated by laminar flow or by turbulent flow. The Peclet number, as the Reynolds number, classifies a characteristic of the flow. This time, though, it is the mass transfer. Peclet number is defined to be the ratio of the rate of advection of a physical quantity by the flow to the rate of diffusion of the same quantity. Therefore, if Pe has a high value, the mass transfer is dominated by the advection of the flow meaning that the mass/particles move according to the bulk motion of the flow. On the other hand, for low Pe, the mass transfer is dominated by diffusion and concentration gradients. 𝑃𝑒𝑐𝑙𝑒𝑡 𝑛𝑢𝑚𝑏𝑒𝑟 = 𝑎𝑑𝑣𝑒𝑐𝑡𝑖𝑣𝑒 𝑡𝑟𝑎𝑛𝑠𝑝𝑜𝑟𝑡 𝑟𝑎𝑡𝑒 𝑑𝑖𝑓𝑓𝑢𝑠𝑖𝑣𝑒 𝑡𝑟𝑎𝑛𝑠𝑝𝑜𝑟𝑡 𝑟𝑎𝑡𝑒 = 𝑈 · 𝐿 𝐷 Peclet number is calculated with a velocity magnitude (U), a characteristic length (L) and a diffusion constant (D). The diffusion constant was fixed and estimated as 4.48·10-11 m2s-1 taking into account that the blood had always the same composition and by considering different proteins and molecules present in the coagulation process [17]. Moreover, a local Peclet number for each location in the cross section was considered in order to obtain meaningful information. Finally, Uij was chosen for the velocity magnitude. Taking all this into account, a low magnitude of the Peclet number can ease side clot due to the enhanced diffusion transport of the coagulation factors in relation to the convection which washes them away. On the other hand, a negative local Peclet number, may also potentially arise the risk of side clot because the flow is pushing the coagulation factors towards the vessel wall accumulating them in the same region. Figure 15. Local Peclet numbers in the cross section. The 6.0 mm bump has a thicker layer of low local Peclet numbers on the sides of the irregularity. Red and blue regions are dominated by advection transport. (1) Patient specific models of clotting 22 3.2.3. Tangential Velocity (Utn) Similar to recirculation, two distinct patterns are observed in the key experiments (see Figure 16). In the 1.5 mm bump, the velocity field vectors are almost perpendicular to vessel walls, conversely, in 6.0 mm geometry, velocity vectors tend to be tangential in the region next to the walls at both sides of the bump. The metric for this physical factor was quantifying the tangential tend of the velocity. Therefore, a ratio of the tangential component of the velocity to the vector module was chosen. 𝑈𝑡𝑛 =𝑈𝑡 𝑈 Hence, by observing these flow patterns and having defined its metrics, geometries with high Utn values -close to 1- have higher risk of generating a side clot than the ones with low Utn values -close to 0- because the flow has not a downstream direction and tends to accumulate all coagulation factors in the same cross section. 3.2.4. Wall Shear Stress (WSS) Shear stress is the component of stress coplanar with the material cross section, it arises from the force vector component parallel to the cross section of the material. In other words and applied to the topic, the liquid shear stress in a tubular vessel is the force per unit area applied to blood within the flowing stream, therefore the wall shear stress is the force per unit area applied by the flowing viscous blood to the vessel surface [18]. In a relatively high flow environment of the arterial circuit, blood is a suspension that approximately behaves as a Newtonian fluid [19]. This means that the shear stress can be represented by the following mathematical formula and creates a profile like Figure 17. 𝜏 = 𝜇 · 𝑑𝑢 𝑑𝑦 Figure 16. i-j plane velocity field (Uij velocities). Two distinct patterns are observed. (2) (3) Patient specific models of clotting 23 Figure 17. Blood flow in a tubular chamber generates a parabolic velocity profile. This results in maximal shear at the wall and maximal velocity at the center of the blood stream (Source: [18]). As mentioned back in the section 2, the effect of wall shear stress has direct implications to the creation of platelet plugs by activation, adhesion and aggregation. Shear induced platelet activation is a function of wall shear stress magnitude, exposure time and the shear stress history of the platelet [20]. Figure 18. Wall shear stress calculated with flow going in +Z direction. Abrupt bumps such as 1.5 mm generate higher levels of wall shear stress but concentrated in a small portion of space. The hypotheses are created comparing the experimental and the computational results. The side clot appeared in the long bumps although they generate a lower wall shear stress. According to that, lower wall shear stress increases the risk of side clot formation. A step forward can be done stating that, as the WSS is closely related to the velocity profile (Figure 17), there is a range of WSS -and velocity profilesin which the platelets are activated but they are not flush out downstream because of the blood flow. The thresholds of this range are not known exactly but 6.0 mm geometry is in. Perhaps it is a simplistic model because it does not take into account the exposure time neither the past history of shear stresses. 3.2.5. Vorticity (Vk) Vorticity is a field describes the local spinning motion (the tendency of something to rotate). It is usually associated with turbulent flow but it can also appear in laminar flows. In Figure 19 it can be observed that 1.5 mm bump generates a high vorticity region next to the sides of the irregularity, while 6.0 mm has a lower vorticity in that region. The last side clot hypothesis can be derived from this point: higher vorticity in Z direction (Vk) next to the Patient specific models of clotting 24 bump reduce the risk of side clot formation. The presence of a vortex may generate secondary flows that can abrade the vessel surface preventing coagulation factors from accumulating and clotting. Figure 19. Vorticity in +Z direction (Vk) at the middle cross section. Vorticity is higher at the bump sides of 1.5 mm bump. Wall shear stress and Vorticity are closely related as both are generated in a similar way. The impact of the flow with an abrupt and high bump creates both, high WSS and high Vk. Therefore the data is tightly correlated and it was decided to work with both variables together as one. 3.3. Clinical variables The clinical variables are the frequently used parameters in the description of an atherosclerotic artery for a clinical diagnosis or patient case classification. Using these structural parameters and knowing the composition of the atherosclerotic lesion, physicians are able to characterize the severity and gravity of the disease and decide which the required treatment is. This project is going to focus only on the geometric clinical variables since are the ones related to hemodynamics. There are several ways to measure these values, each medical technique has its advantages and disadvantages but they can be divided into two generic groups, invasive and non-invasive techniques. Currently, most of the atherosclerotic imaging techniques are invasive as they use catheters. Catheter-based techniques visualize the atherosclerotic plaque locally, which may limit clinical applicability, since it would be impractical and too time consuming to visualize all lesions in all coronary arteries although they usually produce high resolution images. Computed tomography angiography, high-frequency intravascular ultrasounds and optical coherence tomography are examples of invasive measuring methods. On the other hand, magnetic resonance angiography, fluorescein angiography and thermometry are examples of non-invasive techniques, they visualize an extensive area but with lower resolution [15] [16]. Patient specific models of clotting 25 All the geometric clinical variables are measured and computed for a stenotic segment of a vessel:  Vessel area  Minimum area  Average diameter  Minimum diameter  Maximum diameter  Reference Vessel Diameter (rvd)  Total length  Eccentricity  Asymmetry Physician usually only look for high-risk plaques vulnerable to rupture but, as it has been discussed in preceding sections, hemodynamics plays also a critical role in thrombosis and needs to be taken into account. The clinical variables that are being used nowadays are not able to characterize the hemodynamics in the stenotic vessel, they contain few useful information to predict accurately blood flow tendency. Therefore, new clinical variables providing valuable insight need to be created and validated in order to be able to diagnose more precisely and include hemodynamics in the treatment decision. This idea is deeply developed and argued in section 5. Two new clinical variables are added to the previous list in order to overcome the gap of information (see Figure 21):  Slope upstream (Slope A)  Slope downstream (Slope B) Figure 21. Geometric variables used to calculate slopes (Slopei = (e+h)/Li). The slopes contain two dimension information, not only axial dimension but also radial one. Figure 20. Top: Radius used for the calculation of Eccentricity (ecc = 1 – Rmin/Rmax). Bottom: Radius used for the calculation of Asymmetry (asym = 1 – Rmin/Rmax). Patient specific models of clotting 32 4.3. Initial experiments 4.3.1. Geometry selection The first set of ex vivo experiments consists on testing two geometries that satisfy all the hypotheses for every single metric of each pattern, simultaneously. In other words, one geometry satisfying the front clot hypothesis and one geometry satisfying all four side clot hypotheses are selected. Among all the geometries that fulfil these terms, the ones with higher combination of the clotting conditions are chosen. Its values are shown in Table 3. Table 3. Flow values of the control geometries and the two tested geometries. Red values indicate clotting thresholds and green values the presence of clot -they are above the thresholds-. In Table 4 the geometric parameters of the chosen simulations are described. 227, the front clot geometry consists on a high and voluminous bump with a sharp drop, on the other hand, 83 consists on a long thin bump. Two different flow recreations are needed. Table 4. Geometric parameters of geometries 227 and 83. In order to get a sense of the clotting magnitudes, these geometries are run together with 1.5 and 6.0 mm bumps that act as reference or control geometries. All the geometries have been built together and have received the same treatments and coatings. Each experiment consists in three runs of four geometries: two control and two being tested. From run to run some parameters such as spinning time or procoagulant agents’ concentrations may be modified in order to maximize the final signal. Several experiments with these geometries are executed to have enough statistical data. 4.3.2. Results In these initial experiments a higher clotting signal of the two tested geometries with respect to the controls is expected. 227 and 83 are supposed to be, theoretically, the geometries with bigger clots in all the simulated data as they satisfy all clotting hypotheses and have Patient specific models of clotting 33 the most probable conditions of generating thrombus. As can be seen in Figure 26, the obtained results are quite encouraging. Figure 26. Two representative results of the initial experiments of 227 and 83. 227 has to be compared to 1.5 mm while 83 has to be compared to 6.0 mm because of the specific patterns of each one. In almost all runs a clear higher signal is detected in the two tested geometries. In both tested geometries the signal is clearly higher than in the controls, therefore all, or at least some of the hypotheses seem to be trustworthy. These initial results are really promising because they are evident manifestations of the relevance of atherosclerotic lumen geometries and hemodynamics in the thrombus origination. Blood flow can not only stimulate different coagulation pathways and patterns but also enhance these, generating bigger and more dangerous clots. For the quantification of these experiments results see section 4.4.2. All the quantified results are presented together in order to ease the comparison and the understanding of the global results. 4.4. Specific geometries test 4.4.1. Geometry selection Having seen the bright initial results, two new sets of geometries are chosen to keep testing and gaining insights into the influence of flow parameters in coagulation process. Differing from last geometry election, this time, four bumps that generate contradictory hypotheses in side clots are tested. As can be seen in Table 5, there are several simulations that only predict side clot for one or some of the flow parameters, the most representative simulations in each group are the selected. With these bump tests the importance and relevance of every each flow metric or possible interactions might be found. Figure 27. 3D pillars model of the tested geometries. Patient specific models of clotting 34 Table 5. Number of simulations satisfying each contradictory hypotheses and the most representative geometry for each condition. In this selection procedure, following the hypotheses thresholds of the 1.5 and 6.0 mm bumps, the notion of Not True includes the states of Uncertain and False. Moreover, wall shear stress and vorticity metrics are combined because they almost always appear in the same states, there are no cases where one is False and the other True. Yet, the occasions in which one of them is True or False and the other Uncertain, the combined state is Uncertain, so Not True. In Table 6 not only metric values for each elected geometry are presented but also a comparison with the clotting thresholds and the two tested geometries. The selection of the most representative simulation in each condition followed the next characteristics:  48: Side clot is predicted by all metrics except tangential velocity. The simulation with the particular side clot conditions which has the lowest tangential velocity is selected.  230: Side clot is only predicted by Peclet number metric. The simulation with the particular side clot conditions that has the less likely combination of all the metrics to produce side clot but Peclet number is selected.  214: Side clot is only predicted by wall shear stress and vorticity. The simulation with the particular side clot conditions that has the less likely combination of all the metrics to produce side clot but wall shear stress and vorticity is selected.  218: Side clot is predicted by all metrics except wall shear stress and vorticity. The simulation with the particular side clot conditions which has the less likely combined wall shear stress and vorticity to produce side clot is selected. Table 6. Flow metrics of the bumps elected for the specific geometries test. Green values are the geometries already tested, red values indicate the presence of clotting -above clotting thresholds-. Patient specific models of clotting 35 In Table 7 the geometric parameters of the chosen simulations are described. This time, three different flow recreations are needed. Table 7. Geometric parameters of the selected simulations. The testing plan is designed following a True-False clotting implications as seen in Figure 28. The geometries are tested in pairs, together with the control pair resembling the initial experiment. The two couples are grouped in such a way that every experiment provides the maximum information. Therefore, the simulations with more factors with opposite values are tested together in first place (48-230). Figure 28. Guiding charts based on positive and negative metric values. Left: Value table, + if clot is expected, - if it is not and ? if it is uncertain. Right: Decision chart based on True-False decisions taken depending on the appearance -Y- or not -N- of clot in each bump test. It seems that, following the previous chart, some of the hypotheses may be found to be incorrect as the final results in all existing possibilities only consist on one or two flow metrics. These hypotheses discards were initially thought to be possible providing that flow metrics are totally independent and do not interact between them. However, the decision chart was incorrect, not only because flow metrics seem to be related and interact between them, but also because the fact that in two geometries clots were formed having a negative value in a flow metric does not directly imply that in a different geometry that same flow metric cannot be the clotting promoter. However, as the experiments’ order was based on this faulty decision chart, the author has found it necessary to present and explain it. As in the previous experiments, each experiment consists in three runs of four geometries: two control and two being tested. From run to run some parameters such as spinning time Patient specific models of clotting 36 or procoagulant agents’ concentrations may be modified in order to maximize the final signal. Several experiments with these geometries are executed to have enough statistical data. The experiments are started with the pair 48 & 230 and followed by 214 & 218. 4.4.2. Results This type of research is expensive, time consuming and requires large amounts of human blood. Since it is dealing with many unknown parameters, it is crucial to collect as much data and as extensively as possible to make it efficient and useful. Therefore, a database of results is generated. For each experiment, pictures are taken methodically and recorded along with all other relevant data such as agents used, concentrations, superficial treatments, runtime, casted materials, etc. Moreover, some remarks related with the experimentation and the results are added, in addition to new hypotheses, related literature or information and proposals for the next experiments. This methodology is very convenient in giving guidance during research advances, avoiding loss of information and always having well collected and useful data to easily go back to review and analyze. As lots of data are generated, going back to recover specific and individual data without a good classification methodology would be impossible. The analysis and evaluation of these experiment results is more complex than the initial experiments results. In the initial experiments, a tendency from the two geometries to be more prone to generate acuter clots was expected to be detected. However, in these experiments not only the presence of clot needs to be detected but also its intensity, the possible emergence of pattern variations and the flow metrics influence in each clot. Because of this fact, the results section is divided into two subsections: one containing visual and qualitative results and the other embracing the quantitative results analysis. 4.4.2.1. Qualitative results Before the presentation of the results and its following analysis, a common flow loop experiment result needs to be introduced: the full clot. Flow loop experiment can present three types of visual results: no presence of clot, presence of clot with its respective pattern -with higher or lower signal- and the full clot. As its name indicates, a full clot is formed when all the blood in a flow loop coagulates in a single clot. Although this is an undesired result because the creation of a full clot usually signifies no detection of the clotting patterns, it may contain useful information as the high severity of the geometry or that the blood is in a too high hyper-coagulative state and the procoagulant agents’ concentrations or runtime need Figure 29. Full clot extracted from a tested tube. Almost no signal is attached to inner walls as the full clot takes it all away. Patient specific models of clotting 37 to be changed for further runs. Always considering, though, the results obtained with the control geometries. In Figure 29 a full clot and the tube where it is formed can be observed. In the next two figures (Figure 30 and in Figure 31) representative tube results pictures are shown from 48 & 230 experiments and 214 & 218 experiments respectively. Figure 30. Representative experimental results sheet of a test, more precisely 48 & 230. The date of the experiment, some experiment characteristics for each run and the tested tubes can be observed. 48 presents clot in all runs and 230 causes two severe full clots and one small signal. Figure 31. Two representative results of the experiments of 214 & 218. A loss of signal in the control bumps can be detected. Full clot in both tested geometries. Several experiments in order to obtain replicates and enough statistical data were performed. In most of the runs the four tested geometries clotted and, in several occasions, as a full clot. Especially the bumps 218 and 230 which presented the acutest thrombus while controls clotted with a small signal. With all the data, two column graphs are generated in order to represent thoroughly all the visual information collected. Patient specific models of clotting 38 The first plot (Figure 32) is useful to visualize the clotting tendency of each geometry. As it shows, all the simulations tested are prone to coagulate. The control clotting condition is imposed in order to be able to remove all the faulty runs in which in any of the tubes a clot was formed. As all the geometries coagulate, it seems that all flow parameters are relevant, at least partially, of the formation of clots. Some of the clotting threshold may be shifted comprising a bigger range because in half of the simulations the clot was predicted by only a single metric while the other metrics were located in the uncertain region. Figure 32. Percentage of clot presence in each geometry when the control tubes clotted. The fact that some geometries do not clot in all runs it is because of the experimental variability. In the second graph (Figure 33) the clot type proportion can be observed. In two of the geometries the proportion is over 80%, this means that almost every time they clotted it was in a full clot form. Therefore, what in run by run seemed to be severe clots, it is confirmed after all the experimental results recompilation. The bumps of geometries 218 & 230 appear to be dangerous stenotic vessels shape prone to form big thrombus. Figure 33. Proportion of clot types in each geometry between regular pattern signal and full clot presence. Patient specific models of clotting 39 4.4.2.2. Quantitative results In this chapter the quantitative analysis of the initial and the specific geometries experiments is presented. As described in section 4.2.6, the quantification of the clots is measured using Lactate-Dehydrogenase (LDH) and Hemoglobin (Heme) assays. For every tube, both assays are performed. Then, a normalization is done for every experiment run in order to be able to compare results between geometries. The normalization consists on considering a level value of 1 to the geometry with higher levels in each assay and then scaling the other bumps assays levels correspondingly. Therefore, for every run, each tube has a value ranging from 0 to 1 in the two quantitative assays. A way to combine the information of both assays is computing the ratio Heme/LDH. With this ratio, plots in Figure 34 and Figure 35 are generated. Figure 34. Heme/LDH ratio for all the geometries that clotted. The highest ratios, close to 1, appear in 218 & 230 geometries. Figure 34 verifies the las chapter observation about 218 & 230. Its ratio is close to 1 because in most of the runs, these two geometries had the highest levels of Heme and LDH so, after the normalization, its LDH and Heme levels where 1 or close to 1. Therefore, it can be posted that these two bump geometries trend to produce severe clots. An interesting characteristic both atherosclerotic geometries share is the combination of presenting not only flow parameters susceptible to generate side clot but also some flow recirculation. So, new insights about flow parameter interactions are appearing. This combination of both clotting patterns, side and front clot, seems to be a relevant finding for precision medicine in which a deeper investigation could be interesting and useful. New hypotheses and a more exhaustive research about these two geometries will be presented in the following chapters. Patient specific models of clotting 40 Figure 35. Heme/LDH ratio for the geometries that clotted with the expected side and front clot patterns showing a good signal. It discards possible full clots in which the pattern was lost and random patterns. Blue lines display the average for the clot types and two levels are found. A higher level trend for side clots and a lower level trend for front clots. In Figure 35, two level trends can be observed, one for each clot type. Initially, these two trends surprised this author as side clots are usually related with platelets activity. Therefore, if side clots are principally composed by platelets, the side clots ratio should be lower than the front clots ratio as LDH levels should be high in side clots. But as the platelets attach all over the reactive part of the tube, because of the collagen coating, and the front clots are usually less voluminous, so they contain smaller amounts of red blood cells and hemoglobin, front clots ratios decrease considerably triggering these trends. A new quantifying method with more reliable and meaningful information has to be implemented. Few ideas were proposed and some of them tested. In order to eliminate the platelet shift, running tubes without bumps and then subtract its LDH levels and normalize all the other tubes might be a solution. Extracting individually the clots and analyze them one by one might also solve the problem. Finally, cutting the tubes close to the bumps in order to isolate clots and be able to measure them individually was the proposal tested and implemented in the last set of experiments. While the experimentation with the specific geometries was progressing, a slow trend to lose signal of the control geometries was detected. In some of the final flow loops runs in this set of experiments, control geometries had almost no signal. The experimental generation of data needed to be paused. However, enough data had already been recorded. Several attempts to recover the signal are discussed in section 4.6. Patient specific models of clotting 41 4.5. Mapping clotting space and classification of stenotic geometries While the laboratory manufacturing and experimentation was taking place, a way to map the clotting space was being developed. The generation of a map or classification of the clotting space formed by all the simulated stenotic geometries provides new insights in the prediction of coagulation states and in matching performances between geometries. The map contributes an easy way of visualization and classification of clotting patterns. A multivariate analysis and representation research was performed and several computational analysis and classifications were assayed before finding the one that was finally used. As theoretically all geometries behaviors are predicted by the five flow hypotheses parameters, all data information is contained in five variables. Thus, a five dimensions space needs to be mapped. In order to do this, a multivariate analysis called Principal Component Analysis (PCA) was considered the best tool. PCA is a dimension-reduction statistical procedure that can be used to reduce a large set of variables to a small set that still contains most of the information. It uses an orthogonal transformation to convert the initial group of possibly correlated variables into a group of linearly uncorrelated variables called principal components. This transformation is defined in such a way that the first principal component provides the largest possible variance -that is, accounts for as much of the variability in the data as possible-, and each succeeding component has the highest variance possible under the constraint that it is orthogonal to the preceding components. Therefore, using this mathematical procedure, the five dimensional space can be reduced into a three dimensional space that can be visually represented without losing almost any data information. As PCA is sensible to the relative scaling of the original variables, all flow metrics have been normalized, ranging from 0 to 1, before applying the analysis. A PCA was applied to the normalized data and the flow parameters were plotted using the three principal component axis containing more variability. These three principal axis contained, while combined, 95.75% of the data variance. So, almost all the information that flow parameters could provide. Initially, without any post-processing, although the plot consisted on scattered points difficult to interpret, three data clusters could be detected. The first localized geometries were the 1.5 and 6mm bumps. Then, geometries with recirculation were distinguished from the rest in order to differentiate front and side clots -one fifth of the threshold recirculation value was selected- and, finally, the specific tested geometries were also positioned. Figure 36 is the obtained plot. Similar PCA were applied to geometrical metrics and clinical variables of each geometry but no remarkable results were found. Patient specific models of clotting 48 From run to run some parameters such as spinning time or procoagulant agents’ concentrations may be modified in order to maximize the final signal. 4.7.2. Results In the first experiment, geometries 137 and 135 trend to produce a higher signal than 116. The amount of signal was so small that the difference had to be measured quantitatively to find the slight variation. Moreover, in most of the runs the controls did not clot or did not show any pattern, even 230. In the second experiment no good results were obtained. The basic controls 1.5 and 6.0 mm bumps did not clot in the vast majority of the runs. In addition, uncoherent full clots appeared intermittently in some geometries. Not even a hypothesis or a new insight could be found from this experiment. Unfortunately, the bad no-clotting and patternless tendency that could not be fixed before these last experiments despite the several attempts, affected also these results. The fact of working with human blood and biological variables produces an instability and randomness really hard to control. Despite these adverse results, two levels of severity seem to be coexisting inside the severe cluster. The upper geometries with positive Peclet number seem to produce clots less acute than the lower geometries with negative Peclet metrics. The negative Peclet number value pushes all the particles and substances to bumps’ walls. On the other hand, recirculation in the downstream part of the bump accumulates all the substances in a limited region and reduces its velocities. The combination of these two flow metrics suggest really severe clots that can lead to important clinical diagnostics that need an imminent treatment. Patient specific models of clotting 49 5. Computational data analysis 5.1. Overview As this project is in the scope of precision medicine, once the impact of hemodynamics and flow parameters in atherosclerotic vessel geometries is known, several attempts to develop methods including hemodynamics for clinical diagnosis and patient classification are discussed in this chapter. The final aim of this computational data analysis is to be able to predict the different behaviors found in experimental results with the use of the clinical variables. As explained in section 3.3, clinical variables are the information physicians receive and use to generate a diagnosis. Therefore, is it possible to avoid the expensive calculation of flow parameters using clinical or geometric variables and predict the final performance or biophysical mechanisms taking place in an atherosclerotic coronary artery? Are the clinical variables currently used enough to produce satisfactory predictions or new clinical variables need to be introduced? These questions were some of the inquiries this author had before this analysis. The clinical variables were computed from the large geometry data set available in a similar way it is done with the optical computer tomography images (OCT). The bumps were sliced in two dimensional curves as if they were vessel profiles observed by OCT and the clinical variables were calculated from the integration of these profiles. 5.2. Predicting models of flow parameters The initial way to face this diagnosis problem was to create predictive models that from clinical variables were able to predict flow parameters accurately. Several univariate regressions models, starting with linear regression and keep increasing the degree, were generated in order to get some insights of the relations between each flow parameter and each individual clinical variable and be able to create a complex model with the combination of the variables. Observing the performance of these models the small amount of flow description information that clinical variables provided was found. Almost any trend between flow parameters and clinical variables appeared. The coefficient of determination (R2) in every model was really low. For this reason, this author decided to introduce new variables containing useful information for flow predictions: the slopes and the flow magnitude. The slopes can be computed from the same OCTs images, on the other hand, flow requires a more complex method of measuring. However, for achieving a good prediction there is no need for an exact flow measure, with a three level classification -low, medium and highseems to be enough. Luckily, this three level flow classification is easy to obtain and no expensive and sophisticate measure method is needed. Patient specific models of clotting 50 After having added these three new variables, a multivariate regression model with all the clinical variables was created for each flow parameter. In Figure 41 the results of these multivariate model and its analysis are presented. Figure 41. Multivariate regression models for each flow parameter. Top: Four most relevant clinical variables for the prediction of each flow parameter, the order of appearance is from higher to lower variance contribution. Middle: Evolution of R-squared as new variables are added to the regression model. The order of appending is the same found in the top box. Bottom: Flow parameters plotted vs. the two most significant clinical variables. Some trends were found but the prediction models were not good enough. However, the three newly added clinical variables not only improved considerably the models but also were the most relevant variable in all flow parameter models. Meaning that they are useful for flow predictions. Moreover, physical intuitions in how bump geometry affects to flow metrics can be found. For example, the most relevant variables in Peclet number are Slope A and flow. In the 3D plot can be observed how high Slope A tend to generate high Peclet values. From this information, it can be derived that the fluid arriving to a high upstream slope is compressed and then it gains more upward acceleration. On the other hand, if the upstream slope is low, the fluid tends to adapt to wall variation with no upward acceleration gain. These two behaviors are scaled with the flow magnitude as fluid velocity is proportional to it. Patient specific models of clotting 51 The same multivariate regression model was created with the geometric parameters and similar results were obtained. The fact that recirculation, Peclet number and tangential velocity are values based on calculations involving different phases and analysis make them difficult to be predicted. On the other hand, the models obtain better results with wall shear stress and vorticity as their calculation is more straightforward and are highly dependent on flow magnitude. 5.3. Prediction of clotting clusters Having seen the bad predictive potential of the multivariate regressions, a new approach to flow predictions was taken. Instead of attempting to predict the exact value of the flow, different clusters with a defined behavior each are going to be predicted. In this way, there is no need to know the values of the flow parameters as the geometries in each cluster present similar performances. Therefore, if the cluster is known, the clotting characteristics are also known and a precise diagnostic can arise followed by a specific treatment depending on what cluster the atherosclerotic geometry belongs to. Several predicting models are generated using machine learning techniques for cluster matching. Geometric parameters and clinical variables were used as predictors and the cluster was predicted. The cluster partitions are shown in Figure 42. Figure 42. The three data partition ways are plotted. The data was divided up to four clusters. The clusters were -in left plot-: severe cluster in green, front clot cluster in cyan, high Utn central cluster in blue and low Utn central cluster in red. The best predicting models in each data partition were recorded and are exposed in Figure 43. It can be seen that as more clusters to predict, worsts predictions -lower accuracies- and that the integration of flow as a predictor produces a significant increase in all the models’ accuracies. In all these models cross-validation of the results was applied in order to avoid data overfitting. Patient specific models of clotting 52 Figure 43. Accuracy of the best models in each data partition depending on the use of geometric parameters (left) or clinical variables with slopes (right) and the integration of flow as a predictor (red) or no integration (grey). It was found that using both variables collections -geometric and clinicalthe predictions of the severe region had a lot of accuracy, almost 99%. This means that the model is able to predict almost all the severe geometries and, therefore, if a new severe bump appears in the infinite possible atherosclerotic vessel irregularities, this model would be able to classify it as a potential severe clot geometry. Figure 44 shows the performance of the clinical variables model predicting the severe cluster. Figure 44. Performance resume of the Quadratic Support Vector Machine as a model using clinical variables with slopes as predictors. Class 1 refers to severe cluster geometries. In the studied set of data, there are eleven geometries considered causers of severe clots. All of them are well predicted but one -bump 116, already tested as a cluster delimiting geometry- and two extra geometries are considered severe without belonging to the cluster -one of them is geometry 137 that is in the cluster frontier-. It is preferred the model to predict a severe case being wrong than predict a wrong no-severe geometry because of the error price in each prediction, over-treating is always better than do not treat an ill patient. Patient specific models of clotting 53 However, bump slopes were used as predictors in the clinical variables model. A set of extra model were generated without using slopes and they were not able to predict almost any severe geometry. The best two cluster model performance can be observed in Figure 45. It was only able to predict two correct severe geometries. Therefore, in order to be able to diagnose more precisely and include hemodynamics in the treatment decision using these type of cluster prediction methods, new clinical variables providing valuable flow insights need to be created and validated. The created models are not the models that should be used in patient diagnosis. They show the capability of cluster prediction in precision medicine focusing on the atherosclerotic coronary arteries. The fact of being able to classify in groups patients with different bump geometries but same danger or behavior and treatment and then to also be able to predict each group is a huge progress in cardiovascular diseases diagnoses and treatments. Figure 45. Performance of Complex Decision Tree as a model using clinical variables without slopes as predictors. Class 1 refers to severe cluster geometries. Patient specific models of clotting 54 6. Conclusions The objective of this project was to provide some useful insights into the clotting patterns that depend on hemodynamics and the atherosclerotic bumps geometries for cardiovascular precision medicine. This thesis started with a large amount of computational simulations data and with some hypothesis from the observation of two geometries. Thanks to computational fluid dynamics the flow features were identified and the hypothesis arose. However, the biologic experiments were needed because a biophysical process that combines several factors was being studied. The simulation results in solitary may be untrusty, nonetheless, its combination with laboratory experimentation becomes in a great methodology to face the coagulation process understanding. By probing the hypothesis, the influence of hemodynamics in clotting and the relevance of the five flow features being studied were identified. The geometry affects in clotting behavior by changing blood flow parameters. Combinations of side and front clots when hypothesis from both pattern were satisfied appeared, turning into severe clotting. Thanks to the threedimensional representation of the Principal Component Analysis of simulations data a new framework to study the influence of hemodynamics was generated. Moreover, the cluster geometry classification and prediction can be a new approach with high potential to cardiovascular diagnoses. If a classification of atherosclerotic geometries based in their behavior, severity and treatment is generated and a large amount of geometries data is available, both, conjointly, can be very suitable in the CAD treatment. By visualization techniques such as OCT, geometry can be defined and then, using prediction models of clusters a very precise, specific and instantaneous diagnosis is made. This project ends with a lot of work yet to be done but with a clearer and smoother path than before it was started. The first step to be done is restore the flow loop’s signal as soon as possible. The solution seems to come from the inner surface treatments of the reactive sites, as seen, roughness or hydrophobicity control might fix the problem. With the use of the 3D clotting map, isolating parameters and choosing the correct geometry to understand a specific characteristic is easier. However, the introduction of specific drugs within the flow loops is also a possible way to isolate the parameters being studied as they can inhibit some coagulation pathways. Combining these two features with the ex-vivo experiment a better comprehension of the coagulation world is achievable. Once the clotting map is fully understood, filling the empty gaps with new geometries might be another good move in order to be sure to cover all the existing space. Finally, a way to classify all atherosclerotic geometries with a single scale or dimensionless value that provides all the information about its behavior and danger -similar to clusteringwould be highly convenient for cardiovascular clinical diagnoses and treatments. Patient specific models of clotting 55 7. References [1] E. Braunwald [et al.]. Braunwald’s heart disease: a textbook of cardiovascular medicine. Seventh Edition, vol. 2, Philadelphia, Elsevier Saunders, (2005). [2] American Heart Association. Silent Ischemia and Ischemic Heart Disease. Post online, (2015). [3] M. Tomas. Precision medicine at the interface between blood and vessel wall. Master’s Thesis. Massachusetts Institute of Technology, (July 2017). [4] J. M. Balaguer. Mapping the latent variability in patient specific thrombotic responses. Master’s Thesis. Massachusetts Institute of Technology, (April 2016). [5] A. Romeu. Identification of flow-based risk factors for blood coagulation. Master’s Thesis. Massachusetts Institute of Technology, (July 2017). [6] X. Li, P. Fang [et al.]. Mitochondrial reactive oxygen species mediate lysophosphatidylcholine-induced endothelial cell activation. Arteriosclerosis, Thrombosis, and Vascular Biology. 36(6), pp. 1090-1100, (April 2016). [7] C. S. Robbins [et al.]. Local proliferation dominates lesional macrophage accumulation in atherosclerosis. Nature Medicine, 19(9), 1166-1172, (Sept 2013). [8] B. Furie and B. C. Furie. Mechanisms of Thrombus Formation. The New England Journal of Medicine. 359, pp. 938-949, (2008). [9] J. M. Paulus. Platelet size in man. Blood. 46, pp. 321-336, (1975). [10] Y. Ikeda. The role of von Willebrand factor and fibrinogen in platelet aggregation under varying shear stress. The Journal of Clinical Investigation. pp. 1234-1240, (April 1991). [11] C. J. Pallister, M.S. Watson. Haematology. Scion Publishing, pp. 336-347, (2010). [12] D. Bluestein [et al.]. Fluid mechanics of arterial stenosis: Relationship to the development of mural thrombus. Annals of Biomedical Engineering. 25, pp. 344 (March 1997). [13] M. Adel [et al.]. Hemodynamic Shear Stress and Its Role in Atherosclerosis. Jama. 282(21), pp. 2035-2042, (December 1999). [14] A. Javadzadegan [et al.]. Flow recirculation zone length and shear rate are differentially affected by stenosis severity in human coronary arteries. Journal of Physiology, Heart and Circulatory Physiology. 304(4), pp. 559-566, (February 2013). Patient specific models of clotting 56 [15] G. Pasterkamp [et al.]. Techniques characterizing the coronary atherosclerotic plaque: influence on clinical decision making? Journal of the American College of Cardiology. 36, (July 2000). [16] F. Otsuka [et al.]. Clinical classification of plaque morphology in coronary disease. Nature Reviews Cardiology. 11, pp. 379-389, (2014). [17] W. A. Wakeham [et al.]. Diffusion coefficients for protein molecules in blood serum. Atherosclerosis. 25, (December 1976). [18] M. H. Kroll [et al.] Platelets and Shear Stress. Blood, the Journal of the American Society of Hematology. 88, pp. 1525-1541, (September 1996). [19] V. T. Turitto. Blood viscosity, mass transport and thrombogenesis. Prog. Hemost Thromb, 6, pp. 139-177, (1982). [20] J. J. Hathcock. Flow effects on coagulation and thrombosis. Arteriosclerosis, Thrombosis and Vascular Biology. 26, pp. 1729-1737, (July 2006). [21] B.J. Hunt [et al.]. Activation of coagulation and platelets is affected by the hydrophobicity of artificial surfaces. Blood Coagulation & Fibrinolysis: an International Journal in Haemostasis and Thrombosis. 8(4), pp. 223-231, (June 1997). [22] J. M. Sañes. Impact of cholesterol induced hydrophobivity on thrombosis. Master’s Thesis. Massachusetts Institute of Technology, (December 2017). [23] L. Balleisen [et al.]. Comparative investigation on the influence of human bovine collagen types I, II and III on the aggregation of human platelets. Journal of Molecule Medicine. 53(19), pp. 903-905 (October 1975). [24] S. Goto [et al.]. Low concentrations of epinephrine can augment shear stress-induced platelet aggregation. Journal of Cardiology. 24(3), pp. 237-241, (May 1994). [25] S. Butenas [et al.]. Tissue factor activity in whole blood. Blood. 105, pp. 2764-2770, (2005). [26] E. Kathleen [et al.]. Thrombin functions during tissue factor-induced blood coagulation. Blood. 100, pp. 148-152, (2002). [27] Y. Tamada [et al.]. Simple method for platelet counting. Biomaterials. 16(3), pp. 259- 261, (1995). Treball Final del Grau en Enginyeria Física APPENDIX Patient specific models of clotting Albert Ripoll Oliveras supervised by Prof. Kumaran Kolandaivelu co-directed by Prof. Blas Echebarria Boston, May 2018. APPENDIX - Patient specific models of clotting 64 B. Flow loop experiment casting protocol This casting protocol has been fully extracted from the one used in the laboratory written by J.M. Sañe and posted in his work titled Impact of cholesterol induced hydrophobicity on thrombosis in January 2018. Cleaning of the Mold 1. Use the PPE Lift off and tooth brush to thoroughly clean the mold, it acts as a plastic remover. Leave it during 15 or 20 minutes. 2. Rinse with water with the pink brush and the pillars by hand. 3. Wash twice the mold with soap, first with the pink brush and afterwards with the white one. It is important to keep rinsing the brushes each time they are used. 4. Wash the pillars with soap once or twice. You can use the wooden stick if needed. Preparation of the Mold 1. Prepare the work areas with the Blue Pads: The Hood and the Pressure Chamber. 2. Spray twice the mold, mainly the parts that are in contact with the chemicals and the inner sticks with bumps. 3. Spray a wooden stick and use it to lubricate the holes that will contain the pillars. It is vital to avoid breaking them. 4. Put Silly Putty in the top part of the holes that are going to be used. The quantity should be enough to fill half of the hole. Flatten the Silly Putty in the upper part of the holes. 5. With the green needle, do small hole on the Silly Putty to create a planar surface. Avoiding air bubbles is key for the process. 6. Join the 3 parts with screws. Start with the base part by placing the flat screws and leaving them loose, continue with the symmetric parts, placing the star screws. Tighten the screws symmetrically, in order to avoid internal tensions. 7. Make sure that it is perfectly closed, without observable misalignments. If so, unscrew wisely the ill positioned screws and screw the again. Preparation of the Mix Material needed:  Part A chemical  Part B chemical  A 50 ml test tube  Three 10 ml syringes (Two of them are capped with a syringe’s tap and glue) and one 20 ml syringe.  Two wooden sticks for mixing APPENDIX - Patient specific models of clotting 65  Small white cap with a “hose” to fill the mold with the mix. 1. A and B chemical’s ratio is 100/32. For making the Loops, pour 20mL of A and 6,4ml of B into the 20ml and 10ml syringes respectively. Extract the air excess by capping the tap with the finger and suck the syringe slowly to make the bubbles bigger. Once they get bigger, leave the syringe in a vertical position to remove the bubbles easily. 2. Put both chemicals into the 50ml test tube. From now on, all the procedures must be agile and as fast as possible. a) Mix it very well, so it is homogenous. Use the 2 mixing sticks in parallel, moving the recipient to an almost horizontal position and back into a vertical one for several times. At the end, the mix should look brown-ish and dense. Watch out with the bottom part of the test tube, which is the trickiest part to mix. b) Distribute the mix into the two remaining 10ml syringes (capped with glue). The plunger has to be removed in this step. c) Put the two syringes in the centrifuge at maximum speed (around 2300 rpm) for 2:30 min. This will remove the last air bubbles. 3. Once it is centrifuged, there are no bubbles in the mix. Remove the green caps from the syringes with a pliers (first one and once it gets empties, the other one). Put the plunger and the white cap with the “hose” into the syringe. The white cap should be lubricated with the mix during all the mold’s filling. Filling of the Mold 1. Use the first syringe with the white lubricated cap. Inject the mix into the mold slowly. The white cap will get out without any effort. Both syringes will be needed to fill the twelve holes. Never stop to pull the plunger in order to avoid filling the mold with air. 2. The 1.5 mm pillars will need some lubrication by putting them part of the mix inside the pumps. This will avoid annoying air bubbles during the process. 3. Introduce the inner sticks with bumps, make sure they are totally inside. (Good trick: introduce them slowly while spinning). It is recommendable to put the pillars alternately. 4. Introduce the mold into the pressurized chamber and leave it there for 7-8h. Extraction of the mold 1. After 7-8h in the pressurized chamber, extract the mold with the blue pad and leave it in the hood. 2. Unscrew all the screws (flat and stars) helping yourself with the pliers. Start from the base part of the mold and continue with the symmetric parts. 3. Remove all the silly putty from the bottom part of the mold. 4. Cut part of the plastic excess at the upper part of the mold with a cutter. Help yourself with scissors. Then, extract (with cautious) all the pillars with the help of the black Allen wrench. 5. Separate all the parts of the mold, starting from the base. Extract slowly and carefully the plastic tubes. Watch out! The tubes easily deform. APPENDIX - Patient specific models of clotting 66 6. Remove the silly putty that is left inside the reactive part of the Loops with small wooden sticks. Avoid grating the inner surface. 7. With the scissors, remove the extra plastic of the extremes of the tubes. 8. Leave the tubes in the Kuminator oven for 24h (with a whole night it is more than ok) at 50ºC (Kuminator is at 37ºC but it is enough). The next step for the mold is “Cleaning of the Mold” and for the Loops is “Preparation of the Loops”. Preparation of the Loops All the pieces that conform the entire Loop:  3 connectors formed by a middle tube cut (2”)* inside a big tube cut (1”)*.  1 small tube cut (8”)*. Non-reactive part of the loop.  1 tube from the mold. Reactive part of the loop. *Trick: To make sure that the length is the right one, follow the written lines of the pink tape. 1. First of all, to create the connectors insert the middle tub into the big tube. Wetting the tubes would make the insertion much easier as it acts as a lubricant. 2. Glue the connectors with the reactive part of the tubes and dry the joints with UVA rays. Do not put the tubes too close to the UV rays, otherwise the glue will get burned easily. Material needed:  Pink needle  Dymax Glue  Yellow plunger to push up the glue Steps to follow: a) Insert some glue in one of the extremes of the tube and attach the connector. Leave it 30 seconds in the UV rays. It is important avoiding glue inside the reactive part of the tube. b) Repeat the process for the other extreme. c) Last but not least, put some glue in the outer part of the joints to reinforce. 3. The tubes should be cleaned three times. Once with water and twice with soap (10 minutes each phase). 4. Put the small large tube (8”) inside the remaining connector. Spinning the large tube would help to insert it. Help yourself with the tweezers to do it. At the end, the connector should be in the center of the large tube. Before moving to step 5, both parts of the tube, the reactive and non-reactive, should have received all the required treatments for running and experiment in the Kuminator, therefore that means that “Plasma treatment” and “Coating proteins to the tubes before running” go first. 5. Join the two extremes of the large tube to the connectors of the reactive part of the loop. Help yourself with tweezers but be really careful in order to avoid scratches or cracks in the loop. APPENDIX - Patient specific models of clotting 67 Plasma treatment 1. Switch on the power button. 2. Insert the tubes slightly inclined into the chamber. 3. With the air valve closed, switch on the pump button. Wait for two minutes. 4. Open slightly and slowly the electricity valve. Be cautious as doing it too fast would ruin the plasma treatment. The color inside should be a shiny intense pink. Wait for another two minutes. 5. Close the electricity valve, switch off the power and the pump button. Finally, open the air valve. 6. Once the chamber is full of air, remove the tubes from it. In the lab, do the hydrophilic test to the tubes to proof that the treatment has been successful. Coating proteins to the tubes before running 1. PBS: 10% in water, usually put 100ml of PBS into 900ml of water. Do not reuse old PBS. 2. Collagen: 10% in PBS solution, usually 45 0uL of collagen into 4,5ml of PBS solution (enough for 12 tubes). Mix it with a large pipet. 3. Insert the collagen solution inside the reactive part of the loop, which are covered by the plunger of the 1ml syringes in one of the sides. Use the largest pipet tip to fill all the tubes with around 250µl of collagen. Cover the extreme with plastic and leave all the tubes in the fridge. 4. Albumin: 2.5% in PBS solution, usually 30ml of PBS with 750mg of Albumin (is in the fridge) are enough for 12 loops. 5. Insert the albumin solution into the non-reactive part of the loops, with a 20ml syringe. Use 1ml syringes to help take the air out of the loop. Leave it resting with the solution for at least 1 hour before the experiment. 6. Before starting the experiment, empty the non-reactive part of the loops and clean them with the PBS solution. Use the six taped syringes, slowly (1mm/s). 7. Prepare the Tyrode’s solution which would be needed after running the Loops at the Kuminator.  Tyrode: 1 envelope (in the fridge) + 1l of water + 2384mg of Hepes + 80 mg of MgSO4 + 1000mg of NaHCO3. Kuminator Before starting: Note: 2.75ml of blood per loop, therefore the standard amount for a 12 Loop run is 33ml. 1. Prepare ACD 2. Prepare CaCl APPENDIX - Patient specific models of clotting 68 When drawing blood: Blood + ACD Quantity of blood: Vblood [ml] = 3mL x #total loops (33ml per 12 Loops) Quantity of ACD: VACD [ml] = Vblood x 0.1 (3.3ml per 12 Loops) Draw blood: 1. Get the big Falcon tube and introduce 10% ACD 2. Draw blood and introduce it in the recipient with ACD 3. Put the recipient, blood + ACD, at the balancing machine at 37C Preparing loops and pc: Material needed: - 1 green needle per loop - 1 pink needle per run - 1 syringe per run Loops: 1. Introduce the green needle in every loop (air scape) 2. Prepare the syringe with the pink needle PC: 1. Plug machine, green flashing light on the ‘box’ on the oven. 2. Inside oven, check flashing green light. 3. Switch on as to fix the green light of the part inside the oven 4. Open the program Kuminator, and press play icon but not ‘move’ button Blood to loop: Blood + CaCl2 Nrun: qty of loops/run VCaCl [uL] = 3 x Nrun x 67.5 Vblood run [mL] = Nrun x 3 Take blood: 1. Get the small Falcon tube and introduce CaCl2 2. Take out blood from the big Falcon tube with the electric pipet, slow! 3. Introduce the blood fast to the small Falcon, touching the wall of it! 4. Put back the blood to balancing APPENDIX - Patient specific models of clotting 69 Introduce blood to loop: 1. Pour the blood into the syringe, carefully so you do not spill blood when pressing near the edge. 2. Pass the needle under the larger diameter tube (from the non-reactive part) and insert it through the two resting tube walls. 3. Inject the blood carefully by holding the loop with the other hand, guiding the air through the air exit. ** Suggestion: Keep holding the air needle with two fingers in order to take it off as fast as possible not spilling (and wasting) any blood. Running: 1. Closed loops in the white thing 2. To the machine 3. Press move button on PC Post running: Material needed: - 1 thyroid syringe per loop - 10% bleach in big recipient Stopping it: 1. Switch off the part of the machine inside the oven 2. Take out the loops 3. Open loops, blood to bleach, start introducing thyroid to the loops slowly. 4. Task manager of the PC and stop the program.