Continuous monitoring of pain using facial expressions in ICU environment (companion) - our team proposes to develop a facial-expressions-based machine learning algorithm named autobps to predict…

The National Institute of Neurological Disorders and Stroke awarded a $1.365M Cooperative Agreement to Massachusetts General Hospital under the Drug Use and Addiction Research Programs (CFDA 93.279) for the project "Continuous Monitoring of Pain Using Facial Expressions in ICU Environment" (COMPANION). Awarded on August 1, 2025, with a completion date of July 31, 2027, this phased initiative develops AutoBPS, a facial-expression-based machine learning algorithm designed to predict the Electronic Behavioral Pain Score for continuous, automated pain monitoring in intensive care unit (ICU) settings. The UG3 pilot phase involves a single-center observational prospective cohort study (N=48) at MGH to develop the AutoBPS prototype and conduct exploratory analyses examining clinical associations between facial expressions, electroencephalography biosignals, and pain, while meeting FDA regulatory requirements. The subsequent UH3 phase scales the initiative to a multi-center prospective cohort study (N=216) across three academic hospitals—MGH, Brigham and Women's Hospital, and Beth Israel Deaconess Medical Center—to expand, externally validate, and extend the algorithm, examine its utility for evaluating treatment effects as clinical trial endpoints, and assess generalizability to non-ventilated patient populations while implementing bias minimization techniques.

The Research Foundation for the State University of New York at Binghamton serves as a sub-awardee, with Dr. Lijun Yin designated as site principal investigator. Yin's team will design and develop project-specific machine learning and deep neural network models for facial action unit detection and facial expression recognition algorithms. Key deliverables include implementing and evaluating software tools for pain assessment using MGH-collected data, leveraging existing data resources and algorithms from Yin's laboratory for algorithm pretraining and validation, and providing expertise in facial expression data management, neural network algorithm development, and testing. The collaborative effort aims to produce an automated, resource-efficient pain monitoring tool that reduces patient suffering, decreases clinician workload, and enables scalable pain assessment in clinical research and ICU environments.

Continuous monitoring of pain using facial expressions in ICU environment (companion) - our team proposes to develop a facial-expressions-based machine learning algorithm named autobps to predict the electronic behavioral pain score (ebps), a digital endpoint for continuous monitoring of pain using facial expressions in the ICU environment (companion). In the UG3 phase, we will conduct a pilot-phase single-center observational prospective cohort study (n = 48) to develop the autobps machine learning algorithm prototype at MGH. We will perform exploratory analyses to establish clinical associations between facial expressions, EEG biosignals and pain, and to conduct analytical and clinical validations in accordance with the FDA regulatory requirements. In the UH3 phase, we will further scale up to a multi-center prospective cohort study (n = 216) to expand, externally validate, and extend the autobps algorithm at three academic hospitals of MGH, BWH, and bidmc. We will continue to expand the autobps model with newly collected BWH/MGH data and externally validate its performance on bidmc set-aside data. We will extend and examine the utility of autobps for evaluating treatment effects as ICU clinical trial endpoints. Lastly, we will examine its generalizability on non-ventilated patient cohorts and perform bias minimization techniques for model enhancement. We expect that our autobps algorithm will provide an automated, resource-sparing, and reliable assistant tool for continuous pain monitoring in the ICU. Such a tool will help decrease patient suffering, reduce clinician workload, and facilitate scalable automated pain assessment in clinical research.