Neuro-symbolic artificial intelligence for real-time sepsis mortality prediction

Fabrizio De Santis, Gyunam Park, Francesco Zanichelli
April, 2026
PDF Cite DOI

Abstract

Sepsis mortality prediction in intensive care units represents a critical engineering challenge in healthcare systems, where timely intervention through intelligent decision support systems can significantly reduce mortality rates. Current clinical decision support systems rely on black-box machine learning models that lack the interpretability required for safe deployment in life-critical engineering applications. We develop a neuro-symbolic artificial intelligence framework that integrates Logic Tensor Networks with automated clinical knowledge extraction for interpretable mortality prediction. Our novel contributions include: (1) an automated pipeline for transforming clinical guidelines into first-order logic constraints using medical knowledge graphs and rule mining; (2) a dual axiom architecture with weak anchoring that enables context-adaptive learning while maintaining semantic grounding; and (3) a three-tier explainability framework providing concept-level, rule-level, and patient-level interpretability. Using 19,328 sepsis patients from a publicly available real-world clinical dataset, our system achieves 86.45% accuracy at 6-h prediction windows, outperforming traditional machine learning baselines while providing full transparency in decision-making. External validation on 6735 patients demonstrates preserved ranking ability despite domain shift challenges.

Type
Journal article
Publication
Engineering Applications of Artificial Intelligence

Practical problem

Intensive care clinicians need to know early which sepsis patients are likely to die, so they can escalate treatment and allocate scarce resources in time. Bedside scoring systems such as SOFA and APACHE II use only a handful of variables and are only moderately accurate, while machine learning models are more accurate but behave as black boxes, so a physician cannot check the rationale behind a life-or-death prediction. Post-hoc explainers such as SHAP give feature importances rather than medical reasoning and can be misleading when features are correlated. Existing neuro-symbolic attempts rely on hand-written rules, ignore the irregular temporal dynamics of ICU data, and explain only at one level, which keeps them out of clinical use.

Research questions

  1. Can clinical guidelines be turned automatically into first-order logic constraints and embedded in a neural mortality predictor without manual rule engineering?
  2. Does injecting such knowledge improve sepsis mortality prediction over purely data-driven baselines across prediction horizons, and which design choices make the integration work rather than harm?
  3. Does the resulting model expose clinically verifiable reasoning at concept, rule and patient level, and does it transport to an external hospital cohort?

Contributions

  • NeSy-SMP, a neuro-symbolic framework for sepsis mortality prediction that grounds Logic Tensor Networks (LTN) predicates in a bidirectional LSTM over temporal sequences of clinical events plus comorbidity features.
  • An automated knowledge pipeline that builds a SNOMED CT knowledge graph from the Surviving Sepsis Campaign guidelines, mines Horn rules with AnyBURL (confidence above 0.8) and translates them into first-order logic axioms.
  • A dual axiom architecture that separates data axioms from knowledge axioms with a weighted satisfiability loss (0.8/0.2) and weakly anchors each clinical predicate to an established threshold.
  • A three-tier explainability framework covering clinical concept learning verification, medical rule consistency validation and an individual patient reasoning dashboard.
  • An evaluation protocol over 6, 12, 24 and 48-hour lead times on MIMIC-IV with clinical operating-point analysis and external validation on eICU, with code released for reproduction.

Key findings

  • On 19,328 MIMIC-IV sepsis patients with 5-fold cross-validation at threshold 0.5, NeSy-SMP has the highest F1 and recall at every lead time, reaching 86.45% accuracy and 80.35% F1 at 6 hours with AUC 88.33% comparable to XGBoost (88.56%).
  • Ablation shows knowledge axioms alone (LTN-K) lower performance below the base LTN, and weak anchoring with equal axiom weights (LTN-AK) collapses to always predicting death (19.03% F1, 50% recall); only anchoring plus weighting helps.
  • At clinical operating points for a 500-patient-per-year hospital (24-hour window), NeSy-SMP at 85% sensitivity produces 137 false alerts per year versus 190 for BiLSTM, and at 88% specificity catches 63 deaths versus 61 for XGBoost and Random Forest and 55 for BiLSTM.
  • Learned predicates align with clinical thresholds after training (lactate 4.0 mmol/L, bilirubin 2.0 mg/dL), and rule truth values separate deaths from survivors for lactate, bilirubin and white blood cells but not for C-reactive protein or mean arterial pressure.
  • On 6,735 external eICU patients without retraining, all models fall to AUC 62-72% and overestimate risk; tree ensembles keep the highest AUC, NeSy-SMP leads F1 only at 6 hours (61.26%), and rule activation drops sharply (chronic conditions 32.6% to 3.0%), while training costs 706 s versus 64 s for BiLSTM.