Explainable Machine Learning for Sepsis Outcome Prediction Using a Novel Romanian Electronic Health Record Dataset

Andrei-Alexandru Bunea; Ovidiu Ghibea; Dan-Matei Popovici; Ion Daniel; Octavian Andronic

cs.LGApr 6, 2026

Explainable Machine Learning for Sepsis Outcome Prediction Using a Novel Romanian Electronic Health Record Dataset

Authors: Andrei-Alexandru Bunea, Ovidiu Ghibea, Dan-Matei Popovici, Ion Daniel, Octavian Andronic

Organizations: POLITEHNICA Bucharest National University of Science and Technology, Bucharest, Romania · Carol Davila University of Medicine and Pharmacy, Faculty of Medicine, Bucharest, Romania · General Surgery Department, University Emergency Hospital of Bucharest, Bucharest, Romania · Innovation and eHealth Center, Carol Davila University of Medicine and Pharmacy, Bucharest, Romania

Abstract

We develop and analyze explainable machine learning (ML) models for sepsis outcome prediction using a novel Electronic Health Record (EHR) dataset from 12,286 hospitalizations at a large emergency hospital in Romania. The dataset includes demographics, International Classification of Diseases (ICD-10) diagnostics, and 600 types of laboratory tests. This study aims to identify clinically strong predictors while achieving state-of-the-art results across three classification tasks: (1)deceased vs. discharged, (2)deceased vs. recovered, and (3)recovered vs. ameliorated. We trained five ML models to capture complex distributions while preserving clinical interpretability. Experiments explored the trade-off between feature richness and patient coverage, using subsets of the 10--50 most frequent laboratory tests. Model performance was evaluated using accuracy and area under the curve (AUC), and explainability was assessed using SHapley Additive exPlanations (SHAP). The highest performance was obtained for the deceased vs. recovered case study (AUC=0.983, accuracy=0.93). SHAP analysis identified several strong predictors such as cardiovascular comorbidities, urea levels, aspartate aminotransferase, platelet count, and eosinophil percentage. Eosinopenia emerged as a top predictor, highlighting its value as an underutilized marker that is not included in current assessment standards, while the high performance suggests the applicability of these models in clinical settings.

Figures & tables

Figure 1: Descriptive statistics of the dataset: (a) age distribution, (b) sex distribution, and (c) hospitalization duration.

Subset size	Top-10	Top-20	Top-30	Top-40	Top-50
% of patients with full coverage	92.59%	72.40%	45.10%	23.99%	8.68%

Table 1: Coverage of patients from the dataset by subset size of most frequent laboratory tests.

Case study	Variant	Subset	Best Model	Acc.	Prec.	Rec.	F1	AUC
Deceased vs. Disch.	diag	top-40	HistGB	0.843	0.843	0.843	0.843	0.920
Deceased vs. Disch.	w/o diag	top-40	GB	0.794	0.794	0.794	0.794	0.866
Deceased vs. Recov.	diag	top-40	HistGB	0.930	0.932	0.930	0.930	0.983
Deceased vs. Recov.	w/o diag	top-40	SVC / LR	0.880	0.882	0.880	0.880	0.938
Recov. vs. Amelior.	diag	top-20	HistGB	0.780	0.780	0.780	0.780	0.865
Recov. vs. Amelior.	w/o diag	top-40	LR	0.690	0.691	0.690	0.690	0.730

Table 2: Best model results across all three classification tasks.

Figure 2: AUC comparison of individual classifiers for Task 1—Deceased vs. Discharged. Diagnostic feature sets (“diag”) yield higher performance than non-diagnostic ones (“nodiag”), with Gradient Boosting variants achieving the best overall results.

Figure 3: Model performance comparison (AUC) for Deceased vs. Recovered patients across diagnostic subsets.

Figure 4: Model performance comparison (AUC) for Recovered vs. Ameliorated patients across diagnostic subsets. Histogram-based Gradient Boosting models achieve the highest AUC values across all subsets.

Figure 5: SHAP summary plot for Deceased vs. Discharged top classifier (HistGB model trained on the top-40 diagnostic subset). Features are ranked by impact on mortality prediction.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

ICD-10 Prefix	Comorbidity Category	Examples
A, B	Infectious Diseases	A41.9 Sepsis, B20 HIV
C, D	Cancer	C18.9 Colon Cancer, D45 Myeloproliferative Neoplasm
E	Endocrine / Diabetes	E11 Type 2 Diabetes
F	Mental Health	F32 Depressive Episode, F20 Schizophrenia
G	Nervous System	G40 Epilepsy, G20 Parkinson’s Disease
H	Eye / Ear Disorders	H25 Cataract, H66 Otitis Media

Appendix

Table S1 : Mapping of ICD-10 diagnostic code initials to comorbidity categories.

#	Laboratory Test	Count	% of patients
1	Urea	12,216	99.33%
2	Sodium (Na + )	12,214	99.32%
3	Potassium (K + )	12,212	99.30%
4	Glucose	12,209	99.28%
5	AST (Aspartate Aminotransferase)	12,193	99.15%
6	ALT (Alanine Aminotransferase)	12,192	99.14%

Appendix

Table S2 : Distribution of patients for top 50 most frequent laboratory tests.

Subset	Variant	Model	Accuracy	Precision	Recall	F1	AUC
10	diag	SVC	0.8184	0.8184	0.8184	0.8184	0.8981
10	diag	Random Forest	0.8178	0.8178	0.8178	0.8178	0.9000
10	diag	Gradient Boosting	0.8339	0.8339	0.8339	0.8339	0.9155
10	diag	Logistic Regression	0.8051	0.8052	0.8051	0.8051	0.8971
10	diag	HistGB	0.8350	0.8351	0.8350	0.8350	0.9144
10	nodiag	SVC	0.7658	0.7667	0.7658	0.7656	0.8399

Appendix

Table S3 : Performance of all base models for Task 1 (Deceased vs. Discharged).

Subset	Variant	Model	Accuracy	Precision	Recall	F1	AUC
10	diag	SVC	0.8767	0.8768	0.8767	0.8767	0.9556
10	diag	Random Forest	0.8814	0.8824	0.8814	0.8813	0.9560
10	diag	Gradient Boosting	0.8930	0.8931	0.8930	0.8930	0.9619
10	diag	Logistic Regression	0.8674	0.8676	0.8674	0.8674	0.9560
10	diag	HistGB	0.9116	0.9118	0.9116	0.9116	0.9734
10	nodiag	SVC	0.8140	0.8140	0.8140	0.8140	0.8925

Appendix

Table S4 : Performance of all base models for Task 2 (Deceased vs. Recovered).

Subset	Variant	Model	Accuracy	Precision	Recall	F1	AUC
10	diag	SVC	0.7535	0.7626	0.7535	0.7513	0.8122
10	diag	Random Forest	0.7465	0.7517	0.7465	0.7452	0.8216
10	diag	Gradient Boosting	0.7442	0.7521	0.7442	0.7422	0.8241
10	diag	Logistic Regression	0.7581	0.7665	0.7581	0.7562	0.8218
10	diag	HistGB	0.7651	0.7707	0.7651	0.7639	0.8374
10	nodiag	SVC	0.6628	0.6654	0.6628	0.6615	0.7352

Appendix

Table S5 : Performance of all base models for Task 3 (Recovered vs. Ameliorated).

Figure S1 : Distribution of patient outcomes at discharge. Red indicates deceased patients, while gray represents discharged patients, including improved, recovered, and worsened cases.

Figure S2 : Model accuracy comparison for Deceased vs. Discharged patients across diagnostic subsets.

Figure S3 : Model accuracy comparison for Deceased vs. Recovered patients across diagnostic subsets.

Figure S4 : Model accuracy comparison for Recovered vs. Ameliorated patients across diagnostic subsets.

Figure S5 : SHAP summary plot for Task 2—Deceased vs. Recovered.

Figure S6 : SHAP summary plot for Task 3—Recovered vs. Ameliorated.

Explore similar work

A Framework for Early Sepsis Prediction via Self-Supervised (JEPA) and Federated Representation Learning
Jul 18, 2026Umair bin Mansoor, Munaf Rashid, Roomi NaqviSelf-Supervised Pre-Training Joint-Embedding Predictive Architecture
Electrical Engg. Department, DHA Suffa University, Karachi, Pakistan · Salim Habib University, Karachi, Sindh, Pakistan · Intel Foundry Services, Intel, Tempe, Arizona, United States
Explainable Transformer Models for Clinical Prediction Tasks on Structured Electronic Health Records
Aug 20, 2026Jun Ni Du, Lukas Adamek, Maxim Kryukov +4Clinical Outcome Prediction Longitudinal EHR Modeling
Data & Computational Science, R&D, Sanofi Toronto, Canada · Data & Computational Science, R&D, Sanofi Barcelona, Spain · Clinical Real World Evidence, R&D, Sanofi Cambridge, USA +2
What Is Missing in Surgical Risk Stratification and Outcome Prediction: A Scoping Review of End-to-End Machine Learning Approaches
Jul 31, 2026Yizhi Dong, Yuhe Ke, Hairil Rizal Abdullah +4Clinical Outcome Prediction Clinical Risk Prediction
Saw Swee Hock School of Public Health, National University of Singapore, Singapore 117549 · Data Science and Artificial Intelligence Lab, Singapore General Hospital, Singapore 169608 · Duke-NUS Medical School, Singapore 169857 +5

Jul 18, 2026J/K move · Enter open · S save

A Framework for Early Sepsis Prediction via Self-Supervised (JEPA) and Federated Representation Learning

Umair bin Mansoor, Munaf Rashid, Roomi Naqvi

Electrical Engg. Department, DHA Suffa University, Karachi, Pakistan · Salim Habib University, Karachi, Sindh, Pakistan · Intel Foundry Services, Intel, Tempe, Arizona, United States

Self-Supervised Pre-Training Joint-Embedding Predictive Architecture Clinical Prediction Longitudinal EHR Modeling Electronic Health Records Clinical Risk Prediction

Open paper AskPDF arXiv

Early sepsis prediction from electronic health records is challenged by irregular sampling, high missingness, and class imbalance. We systematically compare four modeling paradigms -- self-supervised Joint Embedding Predictive Architecture (JEPA) via masked latent prediction, self-supervised VICReg (variance-invariance-covariance regularization) with two-view augmentation, semi-supervised fine-tuning of a VICReg-pretrained encoder, and supervised Temporal Convolutional Network (TCN) -- alongside raw-feature baselines. All models share a common preprocessing pipeline of hourly binning with forward-fill imputation applied to 7 biomarkers selected via sparsity analysis from the MIMIC-III dataset. Our best model (JEPA + XGBoost + mean pooling) achieves AUPRC 0.636 at the time of onset (H0), approaching the SupMix benchmark (0.667) while using 83% fewer biomarkers. The Tier 1 pipeline -- VICReg pretraining followed by semi-supervised fine-tuning and XGBoost -- achieves AUPRC 0.510 at H0, a 3.1

\times

improvement over the raw-feature baseline (0.165) and a 7.6% improvement over the end-to-end supervised TCN (0.474). Crucially, the fine-tuned VICReg encoder exhibits the most temporally persistent representations, degrading only 16.8% from H0 to H10 compared to 47.5% for supervised TCN and 65.3% for JEPA, demonstrating that self-supervised pretraining with task-aware fine-tuning yields features that are both sharp near onset and robust across prediction horizons.

Abstract

Explore similar work

A Framework for Early Sepsis Prediction via Self-Supervised (JEPA) and Federated Representation Learning

Explainable Transformer Models for Clinical Prediction Tasks on Structured Electronic Health Records

What Is Missing in Surgical Risk Stratification and Outcome Prediction: A Scoping Review of End-to-End Machine Learning Approaches