In-Hospital Stroke Prediction from PPG-Derived Hemodynamic Features

The absence of pre-hospital physiological data in standard clinical datasets fundamentally constrains the early prediction of stroke, as patients typically present only after stroke has occurred, leaving the predictive value of continuous monitoring signals such as photoplethysmography (PPG) unvalidated. In this work, we overcome this limitation by focusing on a rare but clinically critical cohort - patients who suffered stroke during hospitalization while already under continuous monitoring - thereby enabling the first large-scale analysis of pre-stroke PPG waveforms aligned to verified onset times. Using MIMIC-III and MC-MED, we develop an LLM-assisted data mining pipeline to extract precise in-hospital stroke onset timestamps from unstructured clinical notes, followed by physician validation, identifying 176 patients (MIMIC) and 158 patients (MC-MED) with high-quality synchronized pre-onset PPG data, respectively. We then extract hemodynamic features from PPG and employ a ResNet-1D model to predict impending stroke across multiple early-warning horizons. The model achieves F1-scores of 0.7956, 0.8759, and 0.9406 at 4, 5, and 6 hours prior to onset on MIMIC-III, and, without re-tuning, reaches 0.9256, 0.9595, and 0.9888 on MC-MED for the same horizons. These results provide the first empirical evidence from real-world clinical data that PPG contains predictive signatures of stroke several hours before onset, demonstrating that passively acquired physiological signals can support reliable early warning, supporting a shift from post-event stroke recognition to proactive, physiology-based surveillance that may materially improve patient outcomes in routine clinical care.

💡 Research Summary

This paper tackles a long‑standing gap in stroke prediction research: the lack of pre‑stroke physiological data in most electronic health record (EHR) datasets. By focusing on a rare but clinically important cohort—patients who suffered an in‑hospital stroke while already under continuous monitoring—the authors are able to collect high‑resolution photoplethysmography (PPG) waveforms that precede the event. Using the MIMIC‑III and MC‑MED critical‑care databases, they identified 176 and 158 stroke cases, respectively, with synchronized, high‑quality PPG recordings.

A key methodological contribution is the extraction of precise stroke onset timestamps from unstructured clinical notes. The authors built an LLM‑assisted pipeline (using Gemini 3 Pro) that parses free‑text narratives, extracts explicit temporal expressions, and resolves relative time references. Two neurologists validated a stratified sample, achieving 95 % concordance within a ±15‑minute window. This “temporal anchoring” converts noisy narrative data into reliable labels, enabling supervised learning.

From each PPG signal, 74 morphological biomarkers per cardiac cycle and their derivatives were computed via the pyPPG toolbox. To mitigate inter‑patient variability, each metric was transformed into a relative displacement score (F_rel) by subtracting a subject‑specific baseline (mean of a stable pre‑stroke window) and normalizing by that baseline. Statistical filtering (Cohen’s d > 0.05) and multicollinearity removal (Pearson r > 0.80) reduced the feature set to 17 physiologically meaningful indicators.

The predictive model is a one‑dimensional ResNet, which leverages residual connections to capture subtle temporal dynamics over long windows. Training and internal validation on the MIMIC‑III cohort yielded F1 scores of 0.7956 (4 h before onset), 0.8759 ± 0.0105 (5 h), and 0.9406 (6 h). Importantly, the same model applied without any re‑training to the external MC‑MED cohort achieved even higher F1 scores of 0.9256, 0.9595, and 0.9888 for the corresponding horizons, demonstrating strong generalizability across institutions, sensor setups, and patient demographics.

The authors also define a rigorous labeling scheme: the timeline is anchored at t = 0 (stroke onset). The interval from –480 min to –(T_w + Δ_pre) is labeled “normal,” the window from –(T_w – Δ_pre) to –Δ_warning is labeled “warning,” and a buffer zone prevents label leakage. This design ensures that the model learns to predict the imminent “warning” period rather than exploiting artifacts.

Overall, the study provides the first large‑scale empirical evidence that PPG contains predictive signatures of acute cerebrovascular events several hours before clinical manifestation. It demonstrates that a data‑centric pipeline—accurate temporal anchoring, robust feature engineering, and a lightweight deep‑learning architecture—can achieve high‑performance early stroke warning without extensive hyper‑parameter tuning. Limitations include the modest number of stroke cases (due to rarity), potential variability in sensor type and placement, and reliance on expert validation for LLM‑generated timestamps. Future work should expand to multi‑center, multi‑sensor datasets, incorporate additional modalities (ECG, blood pressure, respiration), and move toward a real‑time clinical decision support system that can alert clinicians and potentially improve outcomes for patients at risk of in‑hospital stroke.