A67-17 Physiologically Explainable Waveform-based, AI-driven Detection of Pulmonary Embolism

Backgrounds Intermediate (submassive) and high-risk (massive) pulmonary embolism (PE) are associated with high morbidity and mortality and not easily detected without imaging analysis. We hypothesized machine learning-based physiologic feature analysis could be helpful to identify such groups and allow us to understand underlying physiologic mechanisms. Materials and Methods From the PE registry and intensive care unit (ICU) waveform database, both at the University of Pittsburgh, patients who have both continuous waveforms of single-lead electrocardiography (ECG) and photoplethysmography (PPG) were selected. After pre-processing, physiologic features including time-domain, frequency-domain, and nonlinear time series were extracted. Age and gender matched patients without PE were also randomly selected. Random forest (RF) model was applied to the high-granular features to identify discernible physiologic markers responsible for hemodynamic changes. The performance of the RL model was displayed in receiver operating characteristic (ROC) curve as well as precision recall curve (PRC), using features from 15, 30-minute, 1, 3, 6, 9, and 12 hours of waveform data after ICU admission. SHapley Additive exPlanations (SHAP) was then employed to rank relevant feature contribution at each timepoint. Pertinent feature changes were compared for physiologic explainability. Results After waveform quality analysis, 169 intermediate-risk PE and 28 high-risk PE patients were identified as suitable for analysis. 500 patients without PE were randomly selected. 5 time domain (standard deviation of successive beat-to-beat, standard deviation of normal-to-normal (NN) intervals, root mean square of successive differences, percentage of NN intervals ≥ 20ms and 50ms), 3 frequency domain (low frequency, high frequency, ratio), and 2 time series (sample entropy (SampEN), moving average) features from ECG, as well as 5 morphologic features (acceleration time (ACC), pleth amplitude, pleth amplitude variation (PAV), slope, energy) used from PPG. With multiclass analysis using RF, the overall AUROC was 0.88-0.91 across different time intervals with AUPRC of 0.77-0.82. The SHAP showed decreased ACC and increased PAV were important contributor for intermediate-risk PE, both decreased ACC and PAV led ML model towards high-risk PE. Notably, decreased SampEN with its slow recovery after 9 hours was observed in high-risk PE, likely as thrombolytic treatment and/or surgical interventions were performed. Conclusions Using a RF model, with relevant physiologic waveform features as input data, intermediate-risk and high-risk PEs were differentiable from non-PE. Also contributing dynamic physiologic PPG (ACC, PAV) as well as ECG (SampEN) feature trends to the RF model indicated the hemodynamic instabilities.