Automated Pain Assessment from Facial Expressions using Machine Learning Techniques

Abstract

Cancer is one of the major public health problems of the 21st century, with chronic pain prevalent in a large section of its affected population. Automated Pain Assessments (APAs) from facial expressions, leveraging advancements in the field of Artificial Intelligence(AI) have been proposed as a solution to this ever-growing problem and continue to be an active area of research. However, challenges surround their implementation and validity of the subjective construct of pain that many APAs purport to measure. Furthermore, existing datasets to train AI/ML models cover clinical and experimental pain, with cancer pain remaining largely unaddressed.

This thesis therefore examines how an APA model can be developed and validated to measure specific pain constructs, using which implications for cancer pain applications are derived or hypothesised. Two typical pipelines associated with APAs- Classical Machine Learning and Deep Learning, are evaluated in the study, with construct validity evaluated at each sequential phase. The former involves the extraction of Action Units (AUs) using open-source toolkits, classified by SVMs and the latter employs a CNN-LSTM architecture, capable of self-learning facial representations. The Deep-Learning branch reduced dependencies on inconsistent toolkit outputs and captured the subjective pain experience (measured through patients' self-reported ratings) marginally better than its SVM counterparts. However, both APAs were aligned more closely with observer ratings than with self-reported pain.

Development and validation are therefore inseparable in APA design, with the selection of ground truths and datasets largely determining the accuracy of the models. This matters most for cancer pain, where facial expression is known to be subdued and patient-reported outcomes carry higher clinical weight, as validated on real-world cancer pain data developed as part of the SENSAI dataset. The proposed models and the results serve as preliminary findings for the SENSAI Application/Research Programme. Finally, the empirical results are evaluated from a critical and business-oriented lens, to provide takeaways for cancer pain applications and different stakeholders.