A flexible framework for sepsis prediction: Standardizing data management and imputation in time series using MIMIC-III

SoftwareX 29 (2025) 102063 Available online 30 January 2025 2352-7110/© 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/bync/4.0/). Contents lists available at ScienceDirect SoftwareX journal homepage: www.elsevier.com/locate/softx Original software publication A flexible framework for sepsis prediction: Standardizing data management and imputation in time series using MIMIC-III Javier Solís-García, Jose E. Sánchez-López, Belén Vega-Márquez∗ Isabel A. Nepomuceno-Chamorro Dpto. Lenguajes y Sistemas Informáticos, University of Seville, Av. Reina Mercedes sn, Seville, 41012, Spain ARTICLE INFO Keywords: Artificial intelligence Sepsis prediction Time series MIMIC-III ABSTRACT Sepsis is a life-threatening immune response to infections, leading to organ dysfunction. Despite technological advances, the application of AI in sepsis prediction faces challenges, particularly due to the lack of standardized approaches for data preprocessing and imputation. This work introduces a new framework aimed at simplifying data management, ensuring AI models trained on time series data are both reliable and comprehensive. The framework facilitates the construction, preprocessing, and imputation of the Mimic-III database from PhysioNet, providing a standardized benchmark for future AI research in early sepsis prediction. Current code version v1 Permanent link to code/repository used for this code version For example: https://github.com/ElsevierSoftwareX/SOFTX-D-24-00565 Permanent link to Reproducible Capsule For example: https://doi.org/10.5281/zenodo.13941779 Legal Code License MIT License Code versioning system used Git Software code languages, tools, and services used Python, Docker, PostgresSQL, Streamlit Compilation requirements, operating environments & dependencies NumPy==1.21.6, Pandas==1.4.2, TQDM==4.64.0, GPyTorch==1.6.0, Ubuntu 22.04, Docker==20.20.28, docker-compose==1.29.2, Nvidia Container Toolkit==1.16.1 If available Link to developer documentation/manual Support email for questions [email protected] 1. Motivation and significance Sepsis is an immune system reaction to infections causing organ dysfunction; this condition is life-threatening, widespread, and expensive [1]. Each year, estimated 31.5 million cases of sepsis occur worldwide, 19.4 million classified as severe, and 5.3 million resulting in death, particularly in high-income countries [2]. One of the determining factors in the survival of people with this syndrome is early detection, thus avoiding possible damage to the most important organs [3]. However, due to the multiple symptoms that come from this pathology, there are moments when health professionals struggle to detect sepsis quickly. The latest advances in technology have led to an increase in the use of Artificial Intelligence (AI) techniques in the analysis of health data, which is of great importance and usefulness in the treatment of severe pathologies such as sepsis [4]. The effectiveness of AI methods ∗Corresponding author. E-mail address: [email protected] (Belén Vega-Márquez). depends largely on the quality of the data with which they are worked on. One of the databases most used by researchers for the study of early sepsis prediction using machine learning is MIMIC-III [5]. The large number of parameters monitored in an intensive care unit (ICU) makes the resulting databases, such as MIMIC-III, very susceptible to problems related to data quality. One of the key challenges in working with ICU data is that it is often recorded at irregular intervals, with different variables collected from patients at different times [6,7]. As a result, researchers need to preprocess the data, making decisions about how to label it, which patients to include, and how to match cases to control-onset. When labeling the data, there are various definitions of sepsis available in the literature, and choosing the right one is critical. This decision affects the analysis, affecting both the prevalence of sepsis cases and the https://doi.org/10.1016/j.softx.2025.102063 Received 23 October 2024; Received in revised form 9 January 2025; Accepted 18 January 2025 SoftwareX 29 (2025) 102063 2 J. Solís-García et al. complexity of predicting sepsis, depending on whether it occurs earlier or later in the timeline [8]. Although artificial intelligence holds great promise in the advancement of sepsis care, the application of these technologies is not without challenges. Although the number of papers related to sepsis detection is high, there is currently a lack of standardized approaches for the study, preprocessing, and imputation of data in sepsis research. This inconsistency hampers the ability to compare results between studies and to generalize findings to broader patient populations. To address these challenges related to the comparability and reproducibility of machine learning-driven sepsis prediction [8], we propose a new software designed to unify the study, preprocessing, and imputation processes in the field of early sepsis prediction. With this software framework, data download and imputation are immediate. In addition to providing all the code used, it also provides a simple interface so that healthcare professionals are able to use it without difficulty. Our framework allows to impute with 7 different techniques well known in the literature, such as linear interpolation or zero imputation. The motivation for this software is to provide a benchmark framework for future research in AI for sepsis prediction, previously used in [9,10] and therefore allow comparability between different methodologies, which will enable advances in this field. Finally, our work builds upon previous literature but introduces distinct contributions that address existing limitations. For example: •While [11] does not specifically focus on sepsis and uses only Simple Imputation as described in [12], our work targets sepsis explicitly and provides multiple imputation strategies to handle missing values. •Similarly, [13] introduces standardized tasks using MIMIC-III but includes ‘‘Septicemia’’ as part of a phenotype classification task. It is crucial to note the distinction between ‘‘Septicemia’’ (presence of bacteria in the bloodstream) and ‘‘Sepsis’’ (a broader systemic inflammatory response). Our software focuses specifically on creating a standardized dataset for sepsis, addressing a critical gap identified in [14]. •Lastly, while [15] offers tools for general machine learning applications in healthcare, it does not explicitly target sepsis prediction. By consolidating these strengths, our work fills a critical gap in the literature, enabling the reproducibility and standardization necessary to advance AI for sepsis prediction. In the following sections, we provide detailed descriptions of the data preparation pipeline, the implemented imputation techniques, and the framework’s utility as a benchmark for future research. 2. Software description Our software is designed to facilitate the construction of the MimicIII database from PhysioNet, allowing researchers to access the same data used in numerous studies focused on predicting sepsis onset using a variety of machine learning and deep learning techniques, while also providing the preprocessing code to establish the same dataset in the research of AI models for early sepsis prediction. In addition to building the dataset from the database, it implements several basic imputation techniques for time series data, effectively removing the complexity associated with irregularly sampled time series and transforming them into a manageable imputed tabular dataset. 2.1. Software architecture The core of our software takes the publicly available PhysioNet repository [5] to create a PostgreSQL database by utilizing the CSV files available for download from the official PhysioNet website. The database construction is facilitated through the use of Docker, which offers several key advantages. Docker ensures that the software environment is fully reproducible across different machines and operating systems, allowing researchers to avoid configuration issues and dependency conflicts. By encapsulating the database and preprocessing components in Docker containers, we provide a streamlined, userfriendly experience where complex setups are avoided, enabling seamless deployment of the software. This feature not only enhances reproducibility but also simplifies the setup process for users, who can now execute the entire pipeline with minimal effort, a feature no longer available in the original repository. The data from PhysioNet (csv files) consists of ICU admissions, comprising 58,000 ICU stays of over 45,000 distinct patients recorded between June 2001 and October 2012. These records include a wide array of clinical and laboratory variables. To derive meaningful time series data and detect the presence of sepsis in patients, extensive preprocessing steps are necessary. Our software follows the preprocessing steps outlined in a previous study by Moor et al. [9], and these steps are also facilitated through Docker, ensuring a reproducible and consistent execution of the entire pipeline. The preprocessing involves determining the sepsis label, defining patient inclusion criteria, and performing case-control onset matching. Specifically, the detection of sepsis onset is based on the Sepsis-3 criteria, which identifies patients experiencing an increase of 2 points in the SOFA score. Patients under the age of 15, those who develop sepsis within 7 h of ICU admission, and those recorded using the CareVue system — due to underreported negative lab measurements — are excluded. The software further aligns temporal windows between control and sepsis patients to ensure case-control matching. By using Docker to encapsulate these preprocessing tasks, we simplify the process, enabling researchers to reproduce the analysis in different environments without encountering configuration issues. This approach ensures that the preprocessing pipeline can be consistently applied across various systems with minimal effort. Finally, after preprocessing, the dataset is built with 44 selected variables and 6181 patients. All these steps can be seen graphically in Fig. 1. For further details, see [9,10]. The architecture of our software is depicted in Fig. 2, where it shows how Streamlit assists in generating the containers responsible for building the database and applying preliminary preprocessing. It then handles the generation of the final tabular data using the Python scripts, which are also detailed in Fig. 3. 2.2. Software functionalities The main functionalities of this project can be seen represented in Fig. 2, which shows a breakdown of what each part of this project contains, and what tools it uses for its development. The primary functionalities involve building the dataset from the preprocessing stage to be used by both traditional machine learning techniques and deep learning methods. The software includes a suite of algorithms designed to address the irregularity of time series and handle missing values, specifically through a suite of imputation algorithms. This functionality has as a goal to provide a benchmark framework that can assist researchers in developing new algorithms for early sepsis prediction from ICU (Intensive Care Unit) data. Following the initial preprocessing, the data undergoes a series of additional transformations through scripts that reformat the results into an hourly tabulated structure. This process involves: 1. Aggregation of records within the same time using different techniques. It is possible to choose between mean, median, minimum and maximum values. 2. Following the internationally acknowledged guidelines for sepsis definitions, we labeled the data with Sepsis-III [16,17] definition. Sepsis-III requires the presence of suspected infection and organ dysfunction, identified by a 2-point increase in the SOFA score within a 48 to 24 h window around the infection. SoftwareX 29 (2025) 102063 3 J. Solís-García et al. Fig. 1. Steps to be followed in the preprocessing of data and determination of exclusion criteria. Fig. 2. Diagram illustrating the functionality of our software. The data undergoes several transformations before reaching the final stage. This entire process is supported by a Streamlit interface designed to minimize the need for command-line input and provide assistance when necessary. 3. To ensure that the data is relevant and valid for next studies, we follow the patient inclusion criteria [7,9] : (i) patients with sepsis who developed the condition during their ICU stay; (ii) control patients without a sepsis diagnosis but with suspected infection or organ dysfunction; (iii) exclude patients under the age of 15 and those who developed sepsis withing the first 7 h of their ICU stay; (iv) patients with a minimum of 10 observed measurements. 4. Case-control matching is essential to avoid trivial classification tasks, requiring the assignment of 10 random controls for each case to maintain realistic class balance, with data extracted for 48 h prior to sepsis onset [9,18]. While pre-matching is set as the default option in our framework to ensure consistency with prior studies, the software allows users to disable this feature, enabling flexibility for real-world clinical scenarios where incidence rates are not predefined. 5. Imputation strategies: this software allows you to choose between several well-known imputation algorithms, which are described below. Consequently, the data is transformed from a nested structure of lists within lists, each containing various variables and measurements, into a 3D data tensor with dimensions [patient, time point, variable]. For each patient, a vector indicating the presence of sepsis onset is included, positioned at the final time point. Temporal windows of 48 h preceding the onset are considered. The final dataset consists of 6181 patients. Regarding imputation techniques, the software supports several methods [19], including: 1. Linear Interpolation: Values are imputed through linear interpolation using the surrounding values before and after the missing point. 2. Forward Filling: The missing value is imputed by propagating the last observed point forward. If a missing value appears before any real value is observed, it is imputed as 0. 3. Carry Forward: Similar to forward filling, but if a missing value appears before any actual value, it is set as the mean value of the measure for that patient. 4. Indicator Imputation: Missing values are set to 0, and a binary indicator is used to mark the presence of a missing value, with 0 indicating no missing value in the binary indicator. 5. Zero Imputation: Missing values are directly set to 0. 6. Gaussian Process Adapter: This method allows irregularly sampled time series data to be connected to any black-box classifier that can be learned using gradient descent. 7. NaN Imputation: Missing values are filled with NaNs, enabling users to apply their imputation techniques conveniently. Furthermore, the software generates 2D data formats suitable for use with traditional machine learning techniques, with dimensions [patient, time * variable]. The entire preprocessing pipeline is managed through a series of scripts and commands. To enhance user experience, a Streamlit application has been developed. This application guides users through the execution of scripts, automating some steps and allowing users to select SoftwareX 29 (2025) 102063 4 J. Solís-García et al. Fig. 3. Diagram showing the distribution of our Python files used to execute the final preprocessing code, which transforms the data into a tabular format and performs imputation. Fig. 4. Terminal displaying the command required to launch the Streamlit web interface. imputation methods with a simple click. The final output is a set of CSV files containing all processed data. 3. Illustrative examples To run the library, the user simply needs to open the terminal and execute the command shown in Fig. 4. From that point onward, the entire process is guided through the web interface, and any necessary commands will be provided within it. The interface is shown as depicted in Fig. 5 After generating the data through the web interface, the dataset will include the 44 variables described in Table 1. Additionally, Fig. 6 provides information on the number of positive cases and the amount of missing data in the final dataset when no imputation method is applied. 4. Discussion One of the key design decisions in this study was the inclusion of a pre-matching approach to address the class imbalance inherent in ICU datasets. This method, adopted from [9] and extended in [10], assigns 10 control patients for every sepsis case to maintain a balanced dataset. Pre-matching has proven effective in reducing variability caused by imbalanced datasets, facilitating benchmarking under consistent conditions and ensuring the reproducibility of results across studies. Moreover, evidence in [20] supports these proportions, reporting that the percentage of positive sepsis cases among ICU patients typically ranges between 8% and 18%, closely aligning with our pre-matching setup. However, we also acknowledge that real-world clinical scenarios often involve naturally imbalanced incidence rates. To accommodate this variability, our software provides flexibility by allowing users to disable the pre-matching process, enabling alternative approaches tailored to specific research needs. Another central decision in this study was the use of the MIMICIII database, even though the MIMIC-IV database represents the latest version. This choice was driven by the extensive adoption of MIMIC-III in prior research, as highlighted in [9,10,14], making it an invaluable resource for benchmarking and reproducibility. Moreover, the Metavision data used in MIMIC-III are directly comparable to those in MIMIC-IV, ensuring that findings derived from this work remain relevant in contemporary clinical contexts. SoftwareX 29 (2025) 102063 5 J. Solís-García et al. Fig. 5. Screenshot of the Streamlit interface that assists users throughout the entire data generation process. SoftwareX 29 (2025) 102063 6 J. Solís-García et al. Fig. 6. Pie charts showing the proportion of positive and negative cases, as well as the proportion of missing values in the final dataset. Table 1 Measures selected from MIMIC-III for use. Type Measures Vital Systolic Blood Pressure Tidal Volume Set Diastolic Blood Pressure Tidal Volume Observed Mean Blood Pressure Tidal Volume Spontaneous Respiratory Rate Peak Inspiratory Pressure Heart Rate Total Peep Level SpO2 (Pulse Oximetry) O2 Flow Temperature, Celsius FiO2 (Fraction of Inspired Oxygen) Cardiac Output Laboratory AlbuminPotassium Calcium (free) Lactate pO2 Blood Gas Hematocrit pH Blood Gas Hemoglobin pCO2 Blood Gas Platelet Count SO2 Blood Gas Partial Thromboplastin Time Glucose Prothrombin Time (Quick) Troponin T INR (Standardized Quick) Blood Urea Nitrogen Bands (Immature Neutrophils) White Blood Cells Bicarbonate Creatine Kinase Bilirubin Creatine Kinase MB Creatinine Fibrinogen Chloride Lactate Dehydrogenase Sodium Magnesium Nonetheless, we acknowledge the limitations of not transitioning to MIMIC-IV, which includes updated data and excludes CareVue records. By focusing on MIMIC-III, we potentially miss out on leveraging improvements in data quality and representation introduced in the newer version. Future work will address this limitation by extending the methodology and framework to MIMIC-IV, ensuring continued applicability as the field evolves. This extension will include revisiting preprocessing pipelines, updating imputation strategies as needed, and ensuring compatibility with new variables introduced in the MIMIC-IV schema. 5. Impact Although there are numerous algorithm proposals for predicting early sepsis in ICU, it is very difficult to compare them because they typically do not use the same dataset, even building it from the same database, and often do not provide the source code [9]. For this reason, in this software, we propose a tool that facilitates the first and most essential step: standardizing the dataset. We present an easy-touse software to preprocess one of the most widely and widely used databases in the field, particularly for studies focused on the prediction of sepsis, MIMIC-III [21], which will enable greater reproducibility and comparison between different approaches. MIMIC-III is a freely accessible database that contains data from patients admitted to intensive care units (ICU) at Massachusetts General Hospital between 2001 and 2012. It includes clinical data such as vital signs, laboratory results, administered medications, mechanical ventilation, and other treatments. It also contains records of previous illnesses, sepsis diagnosis based on clinical criteria, and progression of the disease during stay in the ICU. The information in MIMICIII is organized in dozens of tabular files (in CSV format), storing different types of data, such as diagnoses, procedures, medications, laboratory measurements, and vital signs. The richness of MIMIC-III’s data makes it a valuable resource for developing and validating early sepsis prediction algorithms. However, its complexity requires careful preprocessing to extract relevant variables and ensure consistency in the analyzes. This software is particularly valuable because, in the past, PhysioNet allowed the construction of this database using Docker technology. However, due to issues with the original code, this option was removed. Our project reintroduces this capability to researchers by providing the necessary code, once again offering the flexibility of Docker for future research endeavors. The data preprocessing offer in our software involved several key steps and is based on [9,10] among others. 6. Conclusions The primary goal of this work is to enhance reproducibility and comparability across studies, a shortcoming highlighted by the review presented in [14]. For this reason, we propose a framework designed to streamline data management for sepsis prediction using artificial intelligence. By facilitating the construction of the MIMIC-III database and providing flexible imputation techniques, our framework addresses key challenges in data preprocessing and time series imputation. The integration of Docker technology reintroduces a previously unavailable capability, enhancing reproducibility and efficiency for researchers. Although MIMIC-IV is the latest version of the database, our choice to focus on MIMIC-III ensures compatibility with a well-established body of literature, making it a critical resource for benchmarking and reproducibility. The similarities between the Metavision data in MIMICIII and MIMIC-IV ensure that our methodology remains relevant to contemporary research needs. Nevertheless, we acknowledge the value of transitioning to MIMIC-IV and outline plans to extend our framework to this updated version in future work, ensuring its adaptability to evolving datasets. Beyond supporting MIMIC-III, this framework lays the groundwork for broader applications in critical care research. Future enhancements could include the incorporation of advanced imputation techniques and SoftwareX 29 (2025) 102063 7 J. Solís-García et al. the integration of additional critical care datasets, further expanding its utility in AI-driven healthcare studies. By addressing these challenges, our work establishes a foundation for advancing reproducibility and fostering innovation in sepsis prediction, ultimately contributing to improved patient outcomes. Credits The logos incorporated in this article were acquired from the Flaticon website. CRediT authorship contribution statement Javier Solís-García: Writing – original draft, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Jose E. Sánchez-López: Validation, Methodology. Belén VegaMárquez: Writing – review & editing, Writing – original draft, Validation, Supervision, Investigation, Conceptualization. Isabel A. Nepomuceno-Chamorro: Writing – review & editing, Writing – original draft, Supervision, Resources, Project administration, Funding acquisition, Formal analysis, Conceptualization. Declaration of competing interest The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Javier Solis Garcia, Jose E. Sanchez Lopez reports financial support was provided by Spain Ministry of Science and Innovation. Belen Vega Marquez, Isabel A. Nepomuceno Chamorro reports article publishing charges and equipment, drugs, or supplies were provided by Spain Ministry of Science and Innovation. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgments This research has been supported by the grant PID2020-117954RBC22 funded by MICIU/AEI/ 10.13039/501100011033 and by TED2021131311B-C21 funded by MICIU/AEI/10.13039/501100011033. This work has also been supported by WinDIAG Project (2021/C005/ 00144832) funded by the European Union - NextGenerationEU, and the Ministry for Digital Transformation and the Civil Service, Government of Spain. References [1] El-Rashidy Nora, Abuhmed Tamer, Alarabi Louai, El-Bakry Hazem M, Abdelrazek Samir, Ali Farman, et al. Sepsis prediction in intensive care unit based on genetic feature optimization and stacked deep ensemble learning. Neural Comput Appl 2022;1–30. [2] Fleischmann Carolin, Scherag André, Adhikari Neill KJ, Hartog Christiane S, Tsaganos Thomas, Schlattmann Peter, et al. Assessment of global incidence and mortality of hospital-treated sepsis. Current estimates and limitations. Am J Respir Crit Care Med 2016;193(3):259–72. [3] Baniasadi Atefeh, Rezaeirad Sepideh, Zare Habil, Ghassemi Mohammad M. Twostep imputation and AdaBoost-based classification for early prediction of sepsis on imbalanced clinical data. Crit Care Med 2021;49(1):e91–7. [4] Liu Andrew C, Patel Krishna, Vunikili Ramya Dhatri, Johnson Kipp W, Abdu Fahad, Belman Shivani Kamath, et al. Sepsis in the era of datadriven medicine: personalizing risks, diagnoses, treatments and prognoses. Brief Bioinform 2020;21(4):1182–95. [5] Johnson Alistair E W, Stone David J, Celi Leo A, Pollard Tom J. The MIMIC code repository: enabling reproducibility in critical care research. J Am Med Inform Assoc 2018;25(1):32–9. [6] Calvert Jacob S, Price Daniel A, Chettipally Uli K, Barton Christopher W, Feldman Mitchell D, Hoffman Jana L, et al. A computational approach to early sepsis detection. Comput Biol Med 2016;74:69–73. [7] Desautels Thomas, Calvert Jacob, Hoffman Jana, Jay Melissa, Kerem Yaniv, Shieh Lisa, et al. Prediction of sepsis in the intensive care unit with minimal electronic health record data: a machine learning approach. JMIR Med Inform 2016;4(3):e5909. [8] Moor Michael, Rieck Bastian, Horn Max, Jutzeler Catherine R, Borgwardt Karsten. Early prediction of sepsis in the ICU using machine learning: A systematic review. Front Med 2021;8:607952. http://dx.doi.org/ 10.3389/fmed.2021.607952, URL https://www.frontiersin.org/articles/10.3389/ fmed.2021.607952/full. [9] Moor Michael, Horn Max, Rieck Bastian, Roqueiro Damian, Borgwardt Karsten. Early recognition of sepsis with Gaussian process temporal convolutional networks and dynamic time warping. In: Machine learning for healthcare conference. PMLR; 2019, p. 2–26. [10] Solís-García Javier, Vega-Márquez Belén, Nepomuceno Juan A, RiquelmeSantos José C, Nepomuceno-Chamorro Isabel A. Comparing artificial intelligence strategies for early sepsis detection in the ICU: an experimental study. Appl Intell 2023;53(24):30691–705. [11] Wang Shirly, McDermott Matthew BA, Chauhan Geeticka, Ghassemi Marzyeh, Hughes Michael C, Naumann Tristan. Mimic-extract: A data extraction, preprocessing, and representation pipeline for mimic-iii. In: Proceedings of the ACM conference on health, inference, and learning. 2020, p. 222–35. [12] Che Zhengping, Purushotham Sanjay, Cho Kyunghyun, Sontag David, Liu Yan. Recurrent neural networks for multivariate time series with missing values. Sci Rep 2018;8(1):6085. [13] Harutyunyan Hrayr, Khachatrian Hrant, Kale David C, Ver Steeg Greg, Galstyan Aram. Multitask learning and benchmarking with clinical time series data. Sci Data 2019;6(1):96. [14] Moor Michael, Rieck Bastian, Horn Max, Jutzeler Catherine R, Borgwardt Karsten. Early prediction of sepsis in the ICU using machine learning: a systematic review. Front Med 2021;8:607952. [15] Jarrett Daniel, Yoon Jinsung, Bica Ioana, Qian Zhaozhi, Ercole Ari, van der Schaar Mihaela. Clairvoyance: A pipeline toolkit for medical time series. 2023, arXiv preprint arXiv:2310.18688. [16] Singer Mervyn, Deutschman Clifford S, Seymour Christopher Warren, ShankarHari Manu, Annane Djillali, Bauer Michael, et al. The third international consensus definitions for sepsis and septic shock (Sepsis-3). Jama 2016;315(8):801–10. [17] Seymour Christopher W, Liu Vincent X, Iwashyna Theodore J, Brunkhorst Frank M, Rea Thomas D, Scherag André, et al. Assessment of clinical criteria for sepsis: for the third international consensus definitions for sepsis and septic shock (Sepsis-3). Jama 2016;315(8):762–74. [18] Nemati Shamim, Holder Andre, Razmi Fereshteh, Stanley Matthew D, Clifford Gari D, Buchman Timothy G. An interpretable machine learning model for accurate prediction of sepsis in the ICU. Crit Care Med 2018;46(4):547–53. [19] Moor Michael, Horn Max, Bock Christian, Borgwardt Karsten, Rieck Bastian. Path imputation strategies for signature models of irregular time series. 2020, arXiv preprint arXiv:2005.12359. [20] Azkárate I, Sebastián R, Cabarcos E, Choperena G, Pascal M, Salas E. Registro observacional y prospectivo de sepsis grave/shock séptico en un hospital terciario de la provincia de Guipúzcoa. Med Intensiv 2012;36(4):250–6. [21] Johnson Alistair EW, Pollard Tom J, Shen Lu, Lehman Li-wei H, Feng Mengling, Ghassemi Mohammad, et al. MIMIC-III, a freely accessible critical care database. Sci Data 2016;3(1):1–9.