JMIR Publications

ABSTRACT

Background:

Approximately 20% of adults in the United States suffer from chronic pain, making it the most common reason for adults seeking medical care. This widespread issue has far-reaching consequences for society, leading to an estimated annual cost of $560 million in terms of medical expenses, lost productivity, and disability. These negative effects underscore the ongoing public health concern posed by chronic pain. On an individual level, the repercussions of inadequate pain management are profound, encompassing physical, psychological, social, and financial hardships for patients. As the initial step in the journey of pain management, pain assessment assumes a critical role. Traditionally, pain assessment relies on self-reports and observational scales, both of which are inherently subjective, time-consuming, and resource-intensive. Consequently, researchers have pioneered the development of automated pain assessment approaches harnessing the power of machine learning to mitigate these challenges.

Objective:

In pursuit of a comprehensive understanding of the current landscape in automated pain assessment, this paper endeavors to conduct a comprehensive survey and review of the recent strides made in this field. The assessment encompasses critical aspects such as datasets, modalities, and machine learning models, which have all played pivotal roles in shaping the trajectory of automated pain assessment. By delving into these key elements, we aim to shed light on the evolving methodologies and approaches that have emerged in recent years.

Methods:

In this comprehensive survey of recent developments within the field of automated pain assessment, we embark on a multifaceted journey. Our survey begins by examining the datasets. We delve into the meta-information and distinctions inherent to various datasets, gaining valuable insights into their applicability and scope. Our analysis extends to the diverse modalities employed within the field, including facial expression analysis, physiological signals, audio data, pupil size variations, and the multimodal inputs. These modalities offer a rich tapestry of sources for pain assessment. Moreover, we dissect the machine learning models that have become pillars in this domain, review their functionalities, strengths, and limitations, thereby providing a holistic overview of the current state of automated pain assessment methodologies.

Results:

The advancements of the automated pain assessment field has been comprehensively surveyed. These advances have been critically assessed and analyzed from various angles, including datasets, modalities, and machine learning models.

Conclusions:

This paper offers a comprehensive survey of the current state-of-the-art in the domain of machine learning-driven automated pain assessment. It initiates by providing a succinct overview of pain's fundamental mechanisms and the accompanying physiological and behavioral responses. Furthermore, a selection of publicly available datasets frequently employed in this field is introduced, including UNBC-McMaster, BioVid, BP4D-Spontaneous, BP4D+, COPE, the YouTube dataset, X-ITE, EmoPain, and SenseEmotion. Subsequently, the paper delves into the exploration of commonly used machine learning techniques for automated pain assessment, with a particular focus on aspects like modality selection, measurement devices, feature extraction, and classification models. Additionally, recent studies concentrating on audio-based and pupil size-based automated pain assessment are also summarized.