A Population-Informed Particle Filter for Robust Physiological Monitoring Using Low-Information Time-Series Measurements

OBJECTIVE: To present the population-informed particle filter (PIPF), a novel filtering approach that incorporates past experiences with patients into the filtering process to provide reliable beliefs about a new patient's physiological state. METHODS: To derive the PIPF, we formulate the filtering problem as recursive inference on a probabilistic graphical model, which includes representations for the pertinent physiological dynamics and the hierarchical relationship between past and present patient characteristics. Then, we provide an algorithmic solution to the filtering problem using Sequential Monte-Carlo techniques. To demonstrate the merits of the PIPF approach, we apply it to a case study of physiological monitoring for hemodynamic management. RESULTS: The PIPF approach could provide reliable beliefs about the likely values and uncertainties associated with a patient's unmeasured physiological variables (e.g., hematocrit and cardiac output), characteristics (e.g., tendency for atypical behavior), and events (e.g., hemorrhage) given low-information measurements. CONCLUSION: The PIPF shows promise in the presented case study, and may have applications to a wider range of real-time monitoring problems with limited measurements. SIGNIFICANCE: Forming reliable beliefs about a patient's physiological state is an essential aspect of algorithmic decision-making in medical care settings. Hence, the PIPF may serve as a solid basis for designing interpretable and context-aware physiological monitoring, medical decision-support, and closed-loop control algorithms.