Analyzing distinct neural network models for oxygen saturation prediction towards a personalized COPD management

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE Analyzing distinct Neural Network Models for Oxygen Saturation Prediction towards a Personalized COPD Management Heribert Pascual, Xavier Masip Advanced Network Architectures Lab (CRAAX) UPC Barcelona Tech Vilanova i la Geltrú., Spain heribert.pascual, [email protected] Albert Alonso Fundació de Recerca Clínic Barcelona- IDIBAPS Hospital Clínic de Barcelona Barcelona, Spain [email protected] Isabel Blanco Department of Pulmonary Medicine Hospital Clínic de Barcelona Barcelona, Spain [email protected] Chronic Respiratory Diseases (CRDs), including Chronic Obstructive Pulmonary Disease (COPD), pose significant global health challenges. Long-term oxygen therapy (LTOT) has been widely adopted to improve respiratory function and quality of life in CRD patients. However, current oxygen dosing approaches show several limitations, including inconvenience and mismatched oxygen flow rates according to patients' activity. To address these challenges, the design of Patient Specific Modeling (PSM) and the utilization of advanced technologies, such as artificial intelligence (AI), can be integrated towards a personalized oxygen dosing. This study presents a first draft of a conceptual architecture combining cloud and edge computing aimed at optimizing oxygen dosing based on individual patient data. The evaluation of sixty neural network architectures in a small-scale test has shown good results, with smaller networks generally outperforming larger ones. Further research is needed to design new testing procedures and explore optimal parameters for the neural networks, including a large patients’ number. Keywords—Chronic obstructive pulmonary disease COPD, artificial intelligence, machine learning, edge computing, blood oxygen saturation, respiratory disease, personalized care I. INTRODUCTION Chronic health conditions, such as cardiovascular diseases, cancer, chronic respiratory problems, or diabetes, have a significant global impact, being recognized as leading causes of death, greatly affecting both the quality of life and the disability of individuals suffering these conditions [1]. Among this spectrum of diseases, Chronic Respiratory Diseases (CRD), and more particularly Chronic Obstructive Pulmonary Disease (COPD), play a substantial role in contributing to the overall burden of disease worldwide. Only in 2019, the number of deaths attributed to COPD surpassed 3.23 million, the major part of them in low–and middleincome countries [2]. Nowadays, being aware that current projections indicate a substantial rise in the prevalence of COPD by 2050 [3], the design of a comprehensive approach to address this public health problem becomes a must. Long-term oxygen therapy (LTOT) has emerged as a widely accepted treatment for patients suffering from COPD and other chronic respiratory diseases that significantly impair their respiratory function [4]. The primary objective of LTOT is to enhance patients' breathing capacity and alleviate respiratory difficulties by administering the required supplementary amount of oxygen through the use of some oxygen supply device. Clinicians heavily rely on the measurement of blood capillary peripheral oxygen saturation (SpO2), as it is an essential indicator that quantifies the proportion of oxygen-saturated hemoglobin relative to the total hemoglobin, serving as a determinant variable for prescribing the proper dosage of supplemental oxygen [4]. Although LTOT notably enhances patients' quality of life, it also presents some drawbacks, basically pointing out to t the need for patients to carry and remain connected to an oxygen supply device throughout the day, what can be inconvenient and psychologically distressing for them. Moreover, the static supplementary oxygen flow rate prescribed by clinicians often fails to align with patients' real-time oxygen demand, leading to a mismatch that can result in long term adverse health effects. During exertion, inadequate oxygen supply may lead to hypoxemia, characterized by symptoms such as dyspnea, muscle fatigue, and reduced exercise tolerance. Conversely, during rest, excessive oxygen levels may result in hyperoxemia, which can lead to oxygen toxicity and potential damage to the lungs and other organs [5]. To overcome these challenges and enhance the management of oxygen therapy, there is a growing interest in developing intelligent and proactive dosing systems. One promising approach involves integrating Patient-Specific Modelling (PSM) into the oxygen dosing process. PSM aims to create personalized models [6] for each individual patient, taking into account their unique physiological characteristics, activity levels, and specific needs. The concept of PSM [7] revolves around the utilization of comprehensive patient data, encompassing medical history, physiological measurements, and real-time monitoring, to construct highly personalized models for managing diverse diseases. In the present era, the integration of PSM technology yields multifaceted outcomes, being a significant one the generation of digital or holographic mixed reality for tumor resection generated from computed tomography patient data [8]. These advancements have significant implications, ranging from precise measurement of implant sizes to the meticulous planning of surgical procedures. Furthermore, PSM holds promise in assisting clinicians with optimal drug treatment selection. Leveraging the patients' individualized PSM, simulations can be performed to predict outcomes for a wide array of treatment options, thereby approximating the potential efficacy and aiding in the decision-making process regarding the most optimal treatment strategy. © 2023 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes,creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. https://dx.doi.org/10.1109/e-Science58273.2023.10254844 The application of PSM applied to CRDs offers a valuable path to both accurately predict oxygen demands and dynamically adjust the oxygen flow rate accordingly. These models will consider a wide range of factors, like lung function, exercise tolerance, and environmental conditions, among others, that should be analyzed. By continuously analyzing and adapting to changes in patients' physiological states, PSM-based dosing systems can optimize the delivery of oxygen, ensuring patients to receive the appropriate amount necessary to meet their specific needs and mitigate potential health risks [6]. By leveraging PSM, the limitations of current rule-based and reactive approaches may be neglected, thus, tailoring the oxygen dosing strategy to the specific requirements of each patient, enhancing patient care and improving overall treatment outcomes in the context of CRDs. The successful implementation of PSM in oxygen dosing systems claims for the integration of advanced technologies, including artificial intelligence (AI), machine learning, data analytics and cloud/edge computing [9]. These powerful tools enable the analysis of vast quantities of patient data, the identification of relevant patterns and trends, and properly support the development of predictive models capable of anticipating changes in SpO2 levels and optimizing oxygen dosing in real-time. Indeed, the availability of edge computing and cloud computing infrastructure further amplifies the capabilities of PSM systems by facilitating rapid data processing, storage, and computation. This seamless integration with existing healthcare systems will enable remote monitoring and management, enhancing the overall efficiency of patient care and overcome the oxygen titration challenges [10]. Moreover, this integration could benefit all the stakeholders, particularly enhancing patients’ quality of life, reducing patients’ disease exacerbations, and extending the oxygen supply devices autonomy, all in all shaping the current reactive paradigm by a predictive and preventive one, thus drastically reducing the costs for the health systems. To validate the effectiveness of PSM-based dosing systems, extensive research studies should be conducted, providing compelling evidence of the potential benefits of personalized oxygen therapy in improving patient outcomes. For instance, research has demonstrated the efficacy of closed loop oxygen titration utilizing the proportional-integral- derivative (PID) control equation. In a study involving patients with COPD, it was found that this approach significantly reduced dyspnea during walking endurance in patients with COPD receiving LTOT [11]. These findings underscore the critical importance of adopting a proactive approach that considers patient-specific factors and leverages PSM to enhance dosing accuracy and the overall effectiveness of oxygen therapy. In this paper, we present a comprehensive approach that combines PSM with AI in a framework able to generate a predictive and tailored oxygen dosing AI model for patients with chronic respiratory diseases, knowing the fact that usual setpoint-based dynamic titration has demonstrated to be useful in patients [12]. Our proposed system seeks for overcoming the limitations of current approaches by harnessing patientspecific data, employing predictive modeling techniques, and integrating real-time monitoring to dynamically adjust the oxygen dosage based on individual needs and activities. The subsequent sections of this paper delve into a literature review about the basic concepts to understand the context in section II, highlighting the advancements while pinpointing existing research gaps. Subsequently, an overview of our proposed system draft is presented in section III, offering insights into the integration of personalized modeling using AI technologies and the edge-cloud continuum. Furthermore, in section IV, a small-scale test is detailed and the obtained results in section V, allowing to draw conclusions in section VI to assess the feasibility and potential of continuing the development of the proposed architecture. Through these analyses, our objective is to contribute to the existing knowledge and provide a foundation for future research in this domain. II. RELATED WORK In recent years, there has been a significant surge of interest in predictive models within the respiratory field, leading to numerous studies focused on improving prediction accuracy. The early contributions by H. Elmoaqet et al. in 2013 [13] laid the foundation for this field. Since then, considerable progress has been made to enhance the predictive capabilities of these systems. H. Elmoaqet et al. dedicated many years to predictive modeling, employing mathematical methods, but failed to achieve significant progress [14], [15]. However, in their most recent contribution [16] introduced a new approach by incorporating AI into a k-step predictive model, proposing a framework for predicting critical levels in physiological signals and introducing a novel performance metric for validation, resulting in remarkable improvements compared to standard autoregressive models. In another recent study by Sam Ghazal et al. [17], AI techniques were employed to predict SpO2 levels following adjustments to mechanical ventilator settings in an Intensive Care Unit (ICU) setting. The authors utilized an Artificial Neural Network (ANN) based on the back-propagation method, along with a Bootstrap aggregation of complex decision trees, as the classifier. While the average results showed promise, the precision for medium and high severity events, which are of utmost importance, was low. This highlights the critical need for improved data quality and quantity to achieve reliable results using predictive approaches. In recent years, the concept of PSM has garnered considerable attention due to its potential applications in healthcare. For instance, in the field of cardiology, specific PSM models have been developed to be to run in edge devices, as cardio twin that analyses Electrocardiogram data to early detect Ischemic Heart Disease (IHD) [18]. Just as another example, PSM has also been utilized in oncology [8]. These real-world examples highlight the broad potential of PSM in enhancing patient care and decision-making across various medical disciplines. Also, it is important to note that regardless of the technology employed, the IT infrastructure must be able to accommodate the specific computational and storage requirements. The Fog-to-Cloud paradigm [19] serves as the foundation for predictive systems, and Xavi Masip et al. applied this paradigm for the first time to the health field, particularly to LTOT [20]. Their approach utilized patients' context, historical data, and biological signals to proactively predict oxygen dosage requirements. By focusing on computing results at edge devices, while preparing the models on the cloud, just transferring a small data amount, alleviates network load and optimizes resource utilization, demonstrating the feasibility of running AI models on currently edge available technologies. Moreover, giving an important advantage in front of other classical proposals, as the edge device does not need to keep a permanent connection to the cloud. To the best of our knowledge, there is limited research specifically addressing the challenges of PSM and edgecomputed SpO2 or oxygen dosage predictions in the domain of CRDs. The preliminary small-scale test presented in this paper aims to begin bridging this, also providing an early evidence regarding the usefulness of PSM in patients suffering from CRDs, with a particular focus COPD. III. ARCHITECTURE OVERVIEW To create the initial outline of the architecture, the fog-to- cloud paradigm has been considered to allocate each task in the most efficient location in terms of computational resources and data privacy. In the cloud, high computational load tasks, large-scale data storage, and knowledge sharing will take place. In the edge (referred to as the fog in [19]) , represented by a healthcare center such as a hospital, the cloud services will be linked to medical users and patients. Finally, the patient, i.e., the end user, will anonymously communicate with the cloud through a linkage generated with their cloud. Fig. 1 shows a diagram of the initial proposal, deeply detailed next. Within the cloud, six modules can be observed. The first module, AI generation, is responsible for generating specific models for each patient using the data collected from the raw database and the patient's available PSM data. Once this model is generated, it will be stored in a historical repository within the AI repository. The second module, the raw database, is where the patients' raw data will be stored, allowing for the association of data with an identifier without revealing the patient's identity. The third module, the AI model repository, is the location where models generated by patients are stored, linked to the patient's data. These models can serve as the initial model for a newly registered patient who does not have sufficient data to create a fully customized model. The fourth module, the patient database, will store the necessary data for the patient's PSM, linking the raw data and the AI models in the repository. The sixth module, the AI model assessment, will add patients to the AI generation module's queue when their model needs to be regenerated due to changes in their disease status, activity, or physical condition. Finally, the diagnostic assistance module will utilize the stored data to generate knowledge that can assist medical professionals in their diagnosis and treatment through the analysis of the entire dataset. At the edge, a front-end component can be found whose main mission is to ease the interaction between medical professionals and the cloud. Its primary task is to establish a linkage between patients and the whole system, enabling them to communicate with the cloud anonymously. The real patient identity linkage will be stored in the patient identity database. The front-end will also link the registered patients' history to the cloud to keep the patient's PSM up to date in case of any changes. Finally, the patient block is responsible for reading user data and forwarding it to the cloud when a connection is available. Periodically, when a connection is established, the sensing & connectivity module will check if the downloaded model is the latest updated version. If not, it will download the new model to perform oxygen dosage predictions on the same device with limited computing capacity as the oxygen supply, or on a connected device. This functional architecture will enable the application of the patient's PSM to the generation of their oxygen dosage algorithm and disease control. Furthermore, various additional small services such as alerts and remote patient monitoring can be added in this architecture in the future. IV. METHODS The main goal of this section is to determine the feasibility of an AI-based model that could be universally applicable to any patient, named best model architecture search. The final results obtained in the proposed small-scale test, provide a preliminary indication of whether it is worth pursuing further research on the generation of PSMs in the field of CRDs, or if it is more appropriate to focus on the development of a generalized model for all users. A. Data Collection Clearance from the Ethics Committee of the Hospital Clínic de Barcelona (HCB) was obtained prior to conducting the test. The study involved collecting data on SPO2 and pulse rates from four patients. These measurements were recorded second-by-second while the patients performed their daily activities at home. Additionally, a prescribed exercise routine was provided to patients with specific instructions to be performed once a day. Data collection took place over a period of approximately two weeks. B. Data Preparation The size of each raw dataset significantly varied across patients: patient 1 contributed 148,000 data points, patient 2 provided 92,000 elements, patient 3 contributed 97,000, and patient 4 had 94,000. From the raw data obtained from each patient, named to as set R, two datasets {Ds, Dh} are generated with prediction horizon, f=20, a continuous stride of s=1, using all the elements available. Moreover, aimed at improving data quality and to avoid inaccuracies from sensor readings, we assume data arrays containing time continuous data; i.e., timestamp difference between two consecutive data points is greater than 5 seconds. The reasoning behind this is that, within a 5 seconds interval, data does not show remarkable changes and it is considered as continuous; otherwise, a new data array is populated. The window size used in this test was selected quite large to rely more on networks architecture than in data itself, a value of wn=50, for the aforementioned continuous data arrays, an intermediate matrix is generated and, finally, its rows are added to the dataset. It is worth noting the resulting dataset consists of matrix of w+1 columns; that is w consecutive data points, Fig. 1. Functional architecture according to s, and an additional column corresponding to the future data point, according to f. In the case of the HR dataset, the algorithm applied is the same, and is appended first the SpO2 block, then the HR block, and finally the result. Equation (1) represents the computation of each element, mij, of the intermediate matrix, given a continuous data array X, and being xi the i-th data point in X. 𝑚𝑖𝑗 ={ 𝑥𝑖+𝑠∗(𝑗−1) , 𝑗≤𝑤 𝑥𝑖+𝑠∗(𝑗−1)+𝑓 , 𝑗 =𝑤+ 1 ∀𝑖 ∈{1,…,|𝑋|−𝑠∗(𝑤−1)} , 𝑗 ∈{1,…,𝑤 + 1} (1) Algorithm 1 represents the pseudocode for the proposed approach to generate dataset D, given the set of collected data R, window size w, stride s and prediction horizon f. Moreover, r ∈ R, represents a data point tuple, containing, among others, the SpO2 measured value, the HR and its timestamp. In brief, after initializing the dataset D, and the continuous data array X to r0, first data point in R (line 1), the algorithm either generates the data array X with time continuous data points (lines 3-4) or generates a new array by initializing X each time two consecutive data points are not considered continuous in time (line 9). Moreover, for each data array X generated, the corresponding intermediate matrix M is computed in computeIntermediateMatrix(·) according to Equation (1), and its rows are added to dataset D, addRows(·) (lines 6-8 and lines 10-12). It is worth recalling that, in case that the number of data points in X is not enough to cover the window size, w (|X|<w), then no rows are added to the dataset from that continuous data array. Finally, the computed dataset is returned (line 13). ALGORITM 1 Both generated datasets were divided into distinct subsets for training, validation, and testing purposes. This division was consistently applied across all subsequent modeling approaches, ensuring uniformity and comparability between the models. C. Models Generation To analyze the performance of different architectures in addressing the research objectives, thirty distinct models were created. These models incorporated a range of architectural configurations, including fully connected (FC) layers, Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) layers, and other Gated Recurrent Units (GRU). Each one of the thirty models was trained using the two prepared data sets for each patient, using a common callback mechanism that monitored the Mean Absolute Error (MAE) during the training process. In Section IV.D the created LSTM architectures are detailed, in Section IV.E the RNN- based ones, in Section IV.F the LSTM + FC based, and finally, in Section IV.G the GRU-based ones D. Network Architectures In Table I, we provide a detailed breakdown of networks referred to as "decreasing" architectures, where the number of units decreases from inputs to outputs. Similarly, Table II presents the corresponding information for "diamond" architectures, which involve an initial increase in units for certain layers followed by subsequent decreases till a one-unit FC layer. A visual representation of these architecture shapes is shown in Figure 1. TABLE I. DECREASING ARCHITECTURES Name Size Data LSTM-ES Small LSTM [ 64, 32 ] + FC [ 1 ] LSTM-EM Medium LSTM [ 256, 128, 64, 32 ] + FC [ 1 ] LSTM-EL Large LSTM [ 1024, 512, 256, 128, 64, 32 ] +FC [1 ] RNN-ES Small RNN [ 64, 32 ] + FC [ 1 ] RNN-EM Medium RNN [ 256, 128, 64, 32 ] + FC [ 1 ] RNN-EL Large RNN [ 1024, 512, 256, 128, 64, 32 ] +FC [ 1 ] LT+FC-ES Small LSTM [ 64, 32 ] + FC [ 1 ] LT+FC-EM Medium LSTM [ 256, 128 ] + FC [ 64, 32, 1 ] LT+FC-EL Large LSTM [ 1024, 512, 256 ] + FC [ 128, 64, 32, 1 ] GRU-ES Small GRU [ 64, 32 ] + FC [ 1 ] GRU-EM Medium GRU [ 256, 128, 64, 32 ] + FC [ 1 ] GRU-EL Large GRU [ 1024, 512, 256, 128, 64, 32 ] +FC [ 1 ] FC-ES Small FC [ 64, 32, 1 ] FC-EM Medium FC [ 256, 128, 64, 32, 1 ] FC-EL Large FC [ 1024, 512, 256, 128, 1 ] To thoroughly evaluate these network designs, we tested three networks of different sizes for both network shapes. Specifically, we examined a small network with a maximum of 64 neurons in its layers, a medium-sized network with a maximum of 256 neurons, and a large network with a maximum of 1024 neurons. To provide a more user-friendly approach for referencing the architectures, each one is named with a descriptive code based on the shape, size, and type of layers. The shapes are denoted as "decreasing" and "diamond," represented by the letters E and I, respectively. Moreover, the layers are categorized by their size, using the descriptors "small" (S), "medium" (M), and "large" (L). For instance, in Table 1, the code "LSTM-ES" is used in the first column and first row, indicating that the network consists of LSTM layers, a decreasing shape and small size. INPUT: R, w, s, f OUTPUT: D 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: Initialize D and X ← X U {r0} for each <r, r’> ∈ R do if areContinuous(<r, r’>) then X ← X U {r’} else if X ≠ Ø then M ← computeIntermediateMatrix(X, w, s, f) D ← addRows(D, M) X ← {r’} if X ≠ Ø then M ← computeIntermediateMatrix(X, w, s, f) D ← addRows(D, M) return D Fig. 2. Shapes for the generated networks TABLE II. DIAMOND ARCHITECTURES Name Size Data LSTM-IS Small LSTM [ 64, 32 ] + FC [ 1 ] LSTM-IM Medium LSTM [ 256, 128, 64, 32 ] + FC [ 1 ] LSTM-IL Large LSTM [ 1024, 512, 256, 128, 64, 32] + FC [ 1 ] RNN-IS Small RNN [ 64, 32 ] FC [ 1 ] RNN-IM Medium RNN [ 256, 128, 64, 32 ] + FC [ 1 ] RNN-IL Large RNN [ 1024, 512, 256, 128, 64, 32] + FC [ 1 ] LT+FC-IS Small LSTM [ 64, 32 ] + FC [ 1 ] LT+FC-IM Medium LSTM [ 256, 128 ] + FC [ 64, 32, 1 ] LT+FC-IL Large LSTM [ 1024, 512, 256 ] + FC [ 128, 64, 32, 1 ] GRU-IS Small GRU [ 64, 32 ] FC [ 1 ] GRU-IM Medium GRU [ 256, 128, 64, 32 ] FC [ 1 ] GRU-IL Large GRU [ 1024, 512, 256, 128, 64, 32] + FC [ 1 ] FC-IS Small FC [ 64, 32, 1 ] FC-IM Medium FC [ 256, 128, 64, 32, 1 ] FC-IL Large FC [ 1024, 512, 256, 128, 64, 32, 1 ] In the second column of both tables, the size is indicated, following the same naming convention described above. Finally, the network structure is presented in the last column, illustrated as “layerType[ l1, l2, l3 ]+layertype2[ l4, l5, l6, ]”. It is worth noting that the last layer of all the networks is always a fully connected layer that utilizes the ReLU activation function, as depicted in the tables alongside other fully connected layers. V. RESULTS This section presents the sorted average results of the MAE, the Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) obtained from a total of 58 networks, each executed six times, for each patient. Table III displays the average errors for patient 1, while Table IV, Table V, and Table VI provide the average errors for patients 2, 3, and 4, respectively. In order to optimize the analysis of the data and to maximize the accuracy of the results, the "LSTM-EM" networks were excluded due to their poor performance. The tables’ first column represents the network's designated name following the here mentioned convention. The "HR" field indicates whether the network was trained solely with SpO2 data ("N" case) or if it also incorporated HR data ("Y" case). Subsequent columns present the calculated MSE, RMSE, and MAPE values. VI. DISCUSSION Upon reviewing the results in Tables 1, 2, 3, and 4, it becomes evident that there is no specific layer architecture, shape, or size that consistently achieved the best results across all patients. However, it is worth noting that the "FC-ES" architecture outperformed the others in patient 2 and 4. On the other hand, the "GRU-ES" architecture attained the best result in patient 1 and 2. Interestingly, for patient 1, the "GRU-IS" architecture achieved exactly the same error as the "GRU-ES" architecture, making it the best performing model. In general, the "FC" architecture, regardless of shape, performed quite well in the majority of cases, with the exception of the third result for patient 1. Conversely, the "LSTM" architecture appeared to be the weakest performer, consistently ranking at the bottom of all tables. TABLE III. PATIENT 1 RESULTS Name HR MAE MSE RMSE MAPE GRUES N 0.51 0.58 0.76 0.55 GRU-IS N 0.51 0.58 0.76 0.55 FC-IS Y 0.53 0.56 0.75 0.57 FC-IM N 0.54 0.59 0.77 0.59 FC-IL N 0.56 0.59 0.77 0.61 FC-EM Y 0.59 0.63 0.80 0.64 FC-ES Y 0.61 0.67 0.82 0.66 FC-EM N 0.61 0.66 0.81 0.66 FC-ES N 0.61 0.67 0.82 0.66 FC-IS N 0.66 0.71 0.84 0.71 FC-IM Y 0.69 0.75 0.87 0.74 RNNES N 0.72 0.99 0.99 0.77 LT+FCES N 0.73 1.03 1.02 0.79 GRUIM N 0.80 1.37 1.17 0.87 LT+FCEL N 0.97 1.69 1.30 1.05 FC-EL N 1.03 1.46 1.21 1.11 LT+FCIM N 1.06 1.72 1.31 1.14 LT+FCEM N 1.10 1.86 1.36 1.18 FC-IL Y 1.17 1.82 1.35 1.26 RNNIL Y 1.37 3.07 1.75 1.48 LT+FCIL N 1.37 2.85 1.69 1.49 RNNEL Y 1.37 3.07 1.75 1.48 RNN-IS Y 1.37 3.07 1.75 1.48 GRUEM Y 1.37 3.07 1.75 1.48 RNNES Y 1.38 3.08 1.75 1.49 RNNEM Y 1.38 3.09 1.76 1.50 GRUEL N 1.39 3.09 1.76 1.50 RNNEL N 1.39 3.09 1.76 1.50 RNN-IS N 1.39 3.10 1.76 1.50 GRUIL N 1.39 3.10 1.76 1.51 RNNEM N 1.40 3.12 1.77 1.52 GRU-IS Y 1.40 3.12 1.77 1.52 RNNIM Y 1.40 3.12 1.77 1.52 RNNIM N 1.42 3.17 1.78 1.53 LT+FCEL Y 1.42 3.17 1.78 1.54 GRUES Y 1.43 3.22 1.79 1.55 RNNIL N 1.44 3.26 1.81 1.56 GRUEL Y 1.45 3.34 1.83 1.57 GRUIL Y 1.45 3.36 1.83 1.58 LT+FCEM Y 1.45 3.34 1.83 1.57 GRUIM Y 1.48 3.59 1.89 1.61 GRUEM N 1.49 3.63 1.91 1.62 LT+FCIS Y 1.49 3.63 1.90 1.62 LT+FCES Y 1.64 7.78 2.79 1.78 LT+FCIL Y 1.69 4.56 2.14 1.84 LSTMIL Y 1.69 8.03 2.83 1.83 LSTMEL Y 1.70 8.09 2.84 1.84 LSTMIM Y 1.70 8.09 2.84 1.84 LSTMES Y 1.70 8.15 2.86 1.85 LSTMIS Y 1.71 8.17 2.86 1.85 LSTMIS N 1.74 7.50 2.74 1.89 LT+FCIM Y 1.88 5.59 2.36 2.05 LT+FCIS N 1.88 12.85 3.58 2.02 LSTMIL N 1.91 12.56 3.54 2.07 LSTMEL N 1.91 12.59 3.55 2.08 LSTMES N 1.92 12.59 3.55 2.08 LSTMIM N 1.92 12.61 3.55 2.08 FC-EL Y 2.06 4.77 2.18 2.21 TABLE IV. PATIENT 2 RESULTS Name HR MAE MSE RMSE MAPE FC-ES Y 1.24 2.90 1.70 1.41 FC-IM Y 1.27 2.96 1.72 1.45 GRU-IS Y 1.28 3.15 1.77 1.46 FC-ES N 1.30 3.27 1.81 1.49 FC-EL N 1.31 3.24 1.80 1.49 FC-EM Y 1.31 3.17 1.78 1.50 FC-IS Y 1.37 3.37 1.84 1.57 FC-IM N 1.46 3.71 1.93 1.65 FC-IL N 1.50 4.02 2.01 1.69 FC-IL Y 1.64 4.09 2.02 1.87 FC-EM N 1.65 4.51 2.12 1.86 RNN-IS N 1.74 5.77 2.40 1.99 RNNES N 1.81 6.06 2.46 2.08 FC-IS N 1.92 5.47 2.34 2.19 LT+FCIL N 2.02 7.67 2.77 2.29 LT+FCIM N 2.04 7.24 2.69 2.34 LT+FCEL N 2.15 8.38 2.89 2.46 LT+FCEM N 2.25 8.16 2.86 2.59 FC-EL Y 2.79 9.94 3.15 3.17 LT+FCIS Y 3.82 26.20 5.12 4.46 GRUEM Y 3.96 26.31 5.13 4.64 RNN-IS Y 3.96 26.32 5.13 4.64 RNNIM Y 3.97 26.36 5.13 4.64 RNNEL Y 3.97 26.39 5.14 4.65 GRUIM N 3.98 26.45 5.14 4.65 GRUIL Y 3.98 26.46 5.14 4.65 RNNIM N 3.98 26.49 5.15 4.66 RNNIL Y 3.99 26.52 5.15 4.66 GRUEM N 3.99 26.53 5.15 4.66 RNNEM Y 3.99 26.55 5.15 4.66 GRUEL Y 3.99 26.55 5.15 4.66 GRUES N 3.99 26.59 5.16 4.67 GRUIM Y 4.00 26.62 5.16 4.67 RNNEL N 4.00 26.62 5.16 4.67 GRU-IS N 4.00 26.62 5.16 4.67 RNNIL N 4.00 26.62 5.16 4.67 RNNES Y 4.00 26.65 5.16 4.68 GRUES Y 4.01 26.80 5.18 4.69 RNNEM N 4.02 26.84 5.18 4.70 GRUIL N 4.02 26.87 5.18 4.70 GRUEL N 4.04 27.02 5.20 4.72 LT+FCIM Y 4.15 28.15 5.31 4.85 LT+FCEM Y 4.16 28.28 5.32 4.86 LT+FCIS N 4.27 32.69 5.72 4.98 LSTMIS Y 4.31 31.35 5.60 5.02 LSTMES Y 4.33 31.58 5.62 5.05 LT+FCEL Y 4.34 30.24 5.50 5.07 LSTMIL Y 4.35 31.81 5.64 5.08 LSTMEL Y 4.45 32.85 5.73 5.19 LSTMIM Y 4.48 33.14 5.76 5.22 LT+FCES Y 4.53 33.91 5.82 5.29 LT+FCES N 4.56 35.48 5.96 5.31 LSTMIL N 4.64 36.36 6.03 5.40 LSTMIS N 4.67 36.73 6.06 5.44 LT+FCIL Y 4.71 34.44 5.87 5.50 LSTMIM N 4.71 37.16 6.10 5.49 LSTMES N 4.76 37.67 6.14 5.54 LSTMEL N 4.82 38.37 6.19 5.61 TABLE V. PATIENT 3 RESULTS Name HR MAE MSE RMSE MAPE GRUES N 1.32 3.45 1.86 1.54 FC-ES N 1.36 3.62 1.90 1.57 FC-EM Y 1.36 3.54 1.88 1.59 FC-EM N 1.36 3.44 1.85 1.58 FC-ES Y 1.37 3.62 1.90 1.59 GRU-IS Y 1.37 3.71 1.93 1.60 GRU-IS N 1.37 3.67 1.91 1.59 FC-IS Y 1.38 3.71 1.93 1.60 GRUEM N 1.38 3.49 1.87 1.62 FC-EL N 1.39 3.50 1.87 1.60 FC-IM Y 1.42 3.75 1.94 1.64 FC-IM N 1.52 3.97 1.99 1.77 FC-IL N 1.66 4.55 2.13 1.95 FC-IS N 1.71 4.72 2.17 1.95 FC-EL Y 1.95 5.92 2.43 2.22 RNNES N 1.98 7.73 2.78 2.31 RNN-IS N 2.01 8.10 2.85 2.32 LT+FCES N 2.06 8.24 2.87 2.37 LT+FCIM N 2.18 8.60 2.93 2.54 LT+FCEM N 2.54 10.36 3.22 2.97 LT+FCEL N 2.56 12.16 3.49 3.00 FC-IL Y 2.63 9.62 3.10 2.98 LT+FCIL N 2.80 13.58 3.68 3.21 LT+FCIM Y 3.32 20.54 4.53 3.91 RNNIM N 4.69 37.47 6.12 5.60 GRUIM Y 4.71 37.68 6.14 5.62 GRUIM N 4.72 37.81 6.15 5.63 GRUIL N 4.72 37.86 6.15 5.63 LT+FCIL Y 4.73 38.04 6.17 5.65 RNNEM Y 4.74 38.21 6.18 5.66 RNN-IS Y 4.76 38.51 6.21 5.69 GRUIL Y 4.78 38.71 6.22 5.71 RNNEM N 4.78 38.75 6.23 5.71 RNNES Y 4.79 38.94 6.24 5.72 RNNIM Y 4.82 39.33 6.27 5.75 RNNIL N 4.82 39.46 6.28 5.76 GRUEL N 4.83 39.62 6.29 5.78 RNNIL Y 4.83 39.62 6.30 5.78 LT+FCEL Y 4.84 39.72 6.30 5.79 GRUEL Y 4.86 40.10 6.33 5.81 RNNEL Y 4.87 40.21 6.34 5.82 RNNEL N 4.87 40.22 6.34 5.82 GRUES Y 4.87 40.29 6.35 5.83 GRUEM Y 4.87 40.36 6.35 5.83 LT+FCES Y 5.01 43.65 6.61 5.98 LT+FCEM Y 5.02 43.01 6.56 6.02 LSTMES Y 5.03 43.98 6.63 6.01 LSTMIM Y 5.06 44.54 6.67 6.05 LSTMEL Y 5.07 44.76 6.69 6.06 LSTMIL Y 5.09 45.09 6.72 6.09 LSTMIS Y 5.11 45.42 6.74 6.11 LT+FCIS N 5.12 46.40 6.81 6.09 LT+FCIS Y 5.13 45.84 6.77 6.14 LSTMEL N 5.20 47.84 6.92 6.20 LSTMIS N 5.21 47.96 6.93 6.21 LSTMES N 5.21 48.00 6.93 6.21 LSTMIM N 5.22 48.12 6.94 6.22 LSTMIL N 5.22 48.21 6.94 6.23 TABLE VI. PATIENT 4 RESULTS Name HR MAE MSE RMSE MAPE FC-ES Y 0.66 0.91 0.95 0.71 GRU-IS N 0.71 1.02 1.01 0.76 FC-IS Y 0.74 1.01 1.00 0.79 FC-IM N 0.75 1.12 1.06 0.81 FC-IS N 0.80 1.21 1.10 0.86 FC-EM N 0.81 1.14 1.07 0.87 GRU-IS Y 0.81 1.35 1.16 0.88 FC-ES N 0.88 1.35 1.16 0.94 FC-IL Y 0.94 1.39 1.18 1.00 FC-IM Y 0.97 1.52 1.23 1.04 FC-EL N 1.03 1.65 1.28 1.10 LT+FCIL N 1.24 2.83 1.68 1.33 LT+FCIM N 1.35 3.25 1.80 1.46 LT+FCEM N 1.49 3.96 1.99 1.62 FC-EM Y 1.56 3.15 1.77 1.66 FC-IL N 1.63 3.41 1.85 1.74 LSTMES N 1.65 10.98 3.31 1.80 FC-EL Y 1.90 4.49 2.12 2.02 LT+FCIL Y 2.16 9.72 3.12 2.36 RNNIM N 2.18 9.91 3.15 2.38 RNN-IS N 2.18 9.92 3.15 2.38 LT+FCIM Y 2.18 9.92 3.15 2.38 GRUIL N 2.18 9.93 3.15 2.38 RNNIM Y 2.18 9.94 3.15 2.38 GRUEM Y 2.18 9.95 3.15 2.39 GRUIM Y 2.19 9.95 3.15 2.39 RNNES Y 2.19 9.95 3.15 2.39 RNNES N 2.19 9.94 3.15 2.39 RNNIL Y 2.19 9.96 3.15 2.39 RNNEL Y 2.19 9.96 3.15 2.39 RNN-IS Y 2.19 9.96 3.16 2.39 GRUIM N 2.19 9.97 3.16 2.39 RNNEM N 2.19 9.97 3.16 2.39 GRUEL N 2.19 9.97 3.16 2.39 RNNEM Y 2.19 9.98 3.16 2.39 RNNEL N 2.19 9.98 3.16 2.39 GRUEL Y 2.19 9.99 3.16 2.40 GRUES Y 2.20 10.00 3.16 2.40 GRUEM N 2.20 10.01 3.16 2.40 RNNIL N 2.20 10.02 3.16 2.40 GRUES N 2.21 10.05 3.17 2.41 GRUIL Y 2.22 10.13 3.18 2.42 LT+FCEL Y 2.25 10.25 3.20 2.45 LT+FCEL N 2.35 8.09 2.85 2.54 LT+FCIS Y 2.36 14.02 3.74 2.57 LSTMIL Y 2.37 14.05 3.75 2.58 LSTMIM Y 2.37 14.05 3.75 2.58 LSTMEL Y 2.38 14.09 3.75 2.59 LSTMIS Y 2.38 14.12 3.76 2.59 LSTMES Y 2.39 14.18 3.76 2.61 LSTMIS N 2.48 14.58 3.82 2.70 LT+FCES Y 2.54 14.98 3.87 2.76 LT+FCEM Y 2.55 12.04 3.47 2.78 LSTMIM N 2.56 18.19 4.26 2.78 LSTMIL N 2.57 18.26 4.27 2.80 LSTMEL N 2.59 18.32 4.28 2.81 LT+FCIS N 2.67 18.88 4.34 2.90 LT+FCES N 2.69 18.82 4.34 2.92 The highest-ranking "LSTM" architecture was "LSTMES," which placed 17th for patient 4. As for the RNN architectures, they did not perform well, slightly better than the LSTM architectures but still failing to produce good results. This is evident in the significant increase in RMSE values for all RNNs. The LSTM architectures with FC layers exhibited similar performance patterns to the RNN architectures, as their results were consistently in close proximity. From an architectural shape perspective, it is apparent that the decreasing architecture yielded the best results across all four patients. However, this observation is not conclusive, as there are diamond-shaped architectures that closely follow in performance for three out of the four patients. In general, the top-ranking networks in the tables consist of both decreasing and diamond-shaped architectures. When examining the dataset perspective, once again, no definitive conclusions can be drawn. Two of the bestperforming results include HR data, while the other two do not. Similarly, analyzing the best-ranked architectures in the tables does not provide any conclusive insights. Even when considering the shape and other factors together, no solid conclusions can be derived. One variable where a weak conclusion can be drawn is the size of the networks. Generally, smaller networks tend to perform better. Upon adopting a broad perspective, it is observed that in most architectures, the larger-sized networks tended to yield worse results compared to their smaller counterparts. However, it is important to note that this observation is not universally applicable and may vary in certain cases. VII. CONCLUSION As deeply analyzed in in the discussion section, no particular neural network architecture stands out. The conclusions drawn from this small-scale test indicate that certain architectures tend to perform better, typically with moderate sizes. This aligns with the future objective of deploying these models on edge devices. Further research will be necessary, both in the architectures that have shown good results, exploring in detail which parameters may be relevant, and refining the testing procedures conducted in this article. Moreover, it is needed to increase the number of patients, as this small-scale experiment has provided only an indication of the path forward, and solid conclusions cannot be drawn from a sample of four patients. As a future line of work, in addition to obtaining a larger sample size of patients, it is proposed to conduct a new experiment similar to the one carried out here, but with a fixed model architecture and variations in the input data achieved by assigning different parameters to the Algorithm1. It would also be relevant to train a user-specific model and examine the implications of attempting to use it to predict outcomes for another user. The long-term goal, if all the conducted tests consistently support the rationale behind the architecture, will be the development of all the modules presented in this initial draft of the architecture. ACKNOWLEDGMENT This work has been supported by the Spanish Ministry of Science and Innovation under grant PID2021-124463OB-I00 funded by MCIN/AEI/10.13039/501100011033 and by ERDF A way of making Europe, the Catalan Government under contract 2021 SGR 00326 and the Catalan Department of Research and Universities. REFERENCES [1] “Noncommunicable diseases.” https://www.who.int/news-room/factsheets/detail/noncommunicable-diseases (accessed Jun. 14, 2023). [2] “Chronic obstructive pulmonary disease (COPD).” https://www.who.int/news-room/factsheets/detail/chronic-obstructive-pulmonary-disease- (copd) (accessed Jun. 16, 2023). [3] A. Benjafield et al., “An estimate of the European prevalence of COPD in 2050,” European Respiratory Journal, vol. 58, no. suppl 65, p. OA2866, Sep. 2021, doi: 10.1183/13993003.CONGRESS-2021.OA2866. [4] J. Pilcher and R. Beasley, “Acute use of oxygen therapy,” Aust Prescr, vol. 38, no. 3, p. 98, Jun. 2015, doi: 10.18773/AUSTPRESCR.2015.033. [5] R. D. Branson, B. R. H. Robinson, R. D. Branson, and B. R. H. Robinson, “Oxygen: when is more the enemy of good?,” 2010, doi: 10.1007/s00134-010-2034-y. [6] C. A. Taylor and C. A. Figueroa, “Patient-Specific Modeling of Cardiovascular Mechanics,” https://doi.org/10.1146/annurev.bioeng.10.061807.1605 21, vol. 11, pp. 109–134, Jul. 2009, doi: 10.1146/ANNUREV.BIOENG.10.061807.160521. [7] J. Carson, R. Van Loon, and P. Nithiarasu, “Mathematical Techniques for Circulatory Systems,” Encyclopedia of Biomedical Engineering, vol. 1–3, pp. 79–94, Jan. 2019, doi: 10.1016/B978-0-12-801238-3.99982-3. [8] S. Mitani et al., “Case‐specific three‐dimensional hologram with a mixed reality technique for tumor resection in otolaryngology,” Laryngoscope Investig Otolaryngol, vol. 6, no. 3, p. 432, Jun. 2021, doi: 10.1002/LIO2.581. [9] X.Masip, H.Pascual, A. Pieto, A.Alonso, I.Blanco “Method and system for determining the amount of oxygen required by a user with respiratory problems”, EP20382872.8 PCT/EP2021/075075 and USA 17916521 [10] S. Mayoralas-Alises, J. M. Carratalá, and S. Díaz-Lobato, “New Perspectives in Oxygen Therapy Titration: Is Automatic Titration the Future?,” Archivos de Bronconeumologia. Elsevier Doyma, 2019. doi: 10.1016/j.arbr.2019.04.002. [11] R. D. Branson, “Oxygen therapy in copd,” Respir Care, vol. 63, no. 6, pp. 734–748, Jun. 2018, doi: 10.4187/RESPCARE.06312. [12] L. M. Kofod, E. Westerdahl, M. T. Kristensen, B. C. Brocki, T. Ringbaek, and F. Hansen, “Clinical Medicine Effect of Automated Oxygen Titration during Walking on Dyspnea and Endurance in Chronic Hypoxemic Patients with COPD: A Randomized Crossover Trial,” J. Clin. Med, vol. 10, p. 4820, 2021, doi: 10.3390/jcm10214820. [13] H. Elmoaqet, D. M. Tilbury, and S. K. Ramachandran, “Predicting oxygen saturation levels in blood using autoregressive models: A threshold metric for evaluating predictive models,” Proceedings of the American Control Conference, pp. 734–739, 2013, doi: 10.1109/ACC.2013.6579923. [14] H. Elmoaqet, D. M. Tilbury, and S. K. Ramachandran, “Evaluating predictions of critical oxygen desaturation events,” Physiol Meas, vol. 35, no. 4, pp. 639–655, 2014, doi: 10.1088/0967-3334/35/4/639. [15] H. Elmoaqet, D. M. Tilbury, and S. K. Ramachandran, “A novel dynamic model to predict abnormal oxygen desaturations in blood,” IEEE MeMeA 2014 - IEEE International Symposium on Medical Measurements and Applications, Proceedings, 2014, doi: 10.1109/MEMEA.2014.6860062. [16] H. ElMoaqet, D. M. Tilbury, and S. K. Ramachandran, “Multi-Step Ahead Predictions for Critical Levels in Physiological Time Series,” IEEE Trans Cybern, vol. 46, no. 7, pp. 1704–1714, 2016, doi: 10.1109/TCYB.2016.2561974. [17] S. Ghazal, M. Sauthier Id, D. Brossier Id, W. Bouachir, P. A. Jouvet Id, and R. Noumeir, “Using machine learning models to predict oxygen saturation following ventilator support adjustment in critically ill children: A single center pilot study,” 2019, doi: 10.1371/journal.pone.0198921. [18] R. Martinez-Velazquez, R. Gamez, and A. El Saddik, “Cardio Twin: A Digital Twin of the human heart running on the edge,” Medical Measurements and Applications, MeMeA 2019 - Symposium Proceedings, Jun. 2019, doi: 10.1109/MEMEA.2019.8802162. [19] X. Masip-Bruin, E. Marín-Tordera, G. Tashakor, A. Jukan, and G. J. Ren, “Foggy clouds and cloudy fogs: A real need for coordinated management of fog-to-cloud computing systems,” IEEE Wirel Commun, vol. 23, no. 5, pp. 120–128, Oct. 2016, doi: 10.1109/MWC.2016.7721750. [20] X. Masip-Bruin, E. Marin-Tordera, A. Alonso, and J. Garcia, “Fog-to-cloud Computing (F2C): The key technology enabler for dependable e-health services deployment,” 2016 Mediterranean Ad Hoc Networking Workshop, Med-Hoc-Net 2016 - 15th IFIP MEDHOCNET 2016, Aug. 2016, doi: 10.1109/MEDHOCNET.2016.7528425.