Machine learning for pain assessment and management

Abstract

Pain is a subjective distressing experience associated with actual or potential tissue damage with sensory, emotional, cognitive and social components. This work aims to develop automatic methods for quantifying pain intensity from physiological and behavioral metrics, and providing real-time clinically interpretable analgesic dosing recommendations personalized according to each patient' evolving pain and physiological condition. Historically, pain in humans has been measured using subjective self-report scales to determine presence and severity. However, these are problematic metrics for both diagnostic and research purposes. For example, self-report is impossible to obtain in various clinical populations, such as unconscious patients or patients with cognitive impairments.

Further, comparisons between people reporting their pain is difficult to do with confidence, as these metrics are highly subjective, depend on previous history with pain and other cognitive and behavioral factors, and can vary over time. Therefore, while current assessment of pain largely relies on the self-report of an individual, the development of an objective, automatic detection/measure of pain may be useful in many research and clinical applications. Such approaches, if successful may not only detect pain, but may provide for a more rational therapeutic intervention. Hence, the objective of this work was to evaluate the use of physiological and behavioral metrics as markers of pain, and to develop automatic methods for objectively quantifying pain. To do so, we focused on three sensing modalities: facial video, autonomic signals from wearable sensors, and functional near-infrared spectroscopy of the brain cortex.

There is often great variability in how people perceive, experience, and physiologically and behaviorally express pain, hence stemming efforts to build a one-size-fits-all pain recognition system. To address this, we proposed novel state-of-the-art machine learning methods for the personalization of the inference process. This approach results in models tailored specifically for individuals that still account for the broader population data. Finally, in this work, we explored the use of reinforcement learning to aid decision making in the intensive care setting by providing personalized pain management interventions.