An Interpretable MACD–SCDM Framework for Sepsis Risk Modeling from Longitudinal Laboratory Trajectories

Nyatega,  Charles o; Adamu, Mohmmed

An Interpretable MACD–SCDM Framework for Sepsis Risk Modeling from Longitudinal Laboratory Trajectories

Date

2026-06-26

Authors

Nyatega, Charles o

Adamu, Mohmmed

Publisher

Electronics

Abstract

Early identification of sepsis from electronic health records (EHRs) remains challeng- ing because longitudinal laboratory measurements are irregularly sampled, incomplete, and dependent on the observation period. This study presents an interpretable tempo- ral feature-engineering framework integrating Moving Average Convergence Divergence (MACD)-based momentum descriptors with a Signal Correlation Decay Measure (SCDM) representing short-term temporal persistence in longitudinal laboratory trajectories. Using the MIMIC-IV Clinical Database Demo (v2.2), 275 hospital admissions from 100 unique patients were screened. Temporal descriptors were derived from lactate, creatinine, white blood cell count, platelet count, and international normalized ratio. Sepsis status was defined at the hospital-admission level using prespecified ICD-9-CM and ICD-10-CM diag- nosis codes, independently of laboratory predictors. Because reliable diagnosis timestamps were unavailable, the task was formulated as admission-level sepsis risk classification rather than prediction of time-stamped sepsis onset. Features were constructed over cu- mulative 0–12 h, 0–18 h, and 0–24 h observation horizons. Lactate was used exclusively as a predictor. HistGradientBoosting and Logistic Regression were evaluated using re- peated patient-grouped cross-validation. Predictive performance varied across observation horizons. Logistic Regression achieved the highest mean AUROC at 24 h (0.711 ± 0.125), while HistGradientBoosting achieved 0.682 ± 0.139. Feature-set ablation indicated that MACD-derived descriptors provided the most consistent incremental contribution relative to conventional features, although their benefit varied by classifier and observation horizon. Permutation importance further demonstrated horizon-dependent feature relevance, with lactate-, INR-, platelet-, and WBC-derived descriptors among the highest-ranked predic-tors

Description

This article was published by electronics

Keywords

sepsis risk modeling, temporal feature engineering, Moving Average Convergence Divergence, Signal Correlation Decay Measure, interpretable machine learning, electronic