We consider the problem of explaining a class of tractable deep probabilistic models, the Sum-Product Networks (SPNs) and present an algorithm ExSPN to generate explanations. To this effect, we define the notion of a context-specific independence tree(CSI-tree) and present an iterative algorithm that converts an SPN to a CSI-tree. The resulting CSI-tree is both interpretable and explainable to the domain expert. We achieve this by extracting the conditional independencies encoded by the SPN and approximating the local context specified by the structure of the SPN. Our extensive empirical evaluations on synthetic, standard, and real-world clinical data sets demonstrate that the CSI-tree exhibits superior explainability.
Routine Electronic Health Record (EHR) blood pressure charting under-samples dynamic physiology, risking missed hemodynamic instability. This study quantifies how EHR-like down-sampling changes the detection and burden of hypo- and hypertension versus continuous monitoring and articulates the consequences for clinical decision support and machine learning label quality. We retrospectively analyzed 78 ECMO-supported pediatric patients (2019–2023). The continuous mean arterial pressure (MAP) captured every 5 s was resampled at intervals from 5 s to 1 h. We screened for 3 min windows of hypotension or hypertension at 10th/90th age-normed thresholds, comparing the per-patient event frequency and burden with EHR-derived recordings. At 10th/90th thresholds, hypotension events fell from 13,936 (5 s) to 3803 (15 min; −72.7%); the EHR captured 3471. Hypertension events dropped from 1573 to 410 (−73.9%); the EHR registered 1587. The EHR data overstated hypertension burden, indicating preferential documentation during prolonged instability while missing brief excursions. Standard EHR sampling significantly under-reports blood pressure derangements in pediatric ECMO. This underreporting of brief events may limit the accuracy of clinical decision support tools and machine learning algorithms in high-acuity patients. High-frequency data acquisition improves event capture and should be prioritized for clinical decision support and machine learning development.
11 April 2025 in Proceedings of the AAAI Conference on Artificial Intelligence
Probabilistic Circuits (PCs) have emerged as an efficient framework for representing and learning complex probability distributions. Nevertheless, the existing body of research on PCs predominantly concentrates on data-driven parameter learning, often neglecting the potential of knowledge-intensive learning, a particular issue in data-scarce/knowledge-rich domains such as healthcare. To bridge this gap, we propose a novel unified framework that can systematically integrate diverse domain knowledge into the parameter learning process of PCs. Experiments on several benchmarks as well as real world datasets show that our proposed framework can both effectively and efficiently leverage domain knowledge to achieve superior performance compared to purely data-driven learning approaches.