Send message to Author

Paper short abstract

Automated pain assessment systems use machine learning to infer pain from behavioural and biometric proxies. Based on an ethnography of APA, this paper shows how chronic pain emerges as a constitutive problem for this field, revealing the data practices that sustain the promises of computable pain.

Paper long abstract

Recent advances in machine learning have produced a striking promise in digital health: that pain, long treated as an irreducibly subjective experience, might be detected and quantified through biometric proxies. Automated pain assessment (APA) systems analyse facial micro-expressions, behavioural cues, and physiological signals with machine learning models to generate algorithmic pain indicators. These systems promise objective, data-driven pain measurement and form part of a growing frontier of automated pain management (Lipp and Hilgartner 2025).

This paper examines how this promise is assembled in practice despite the absence of a stable ground truth for pain. Clinical pain assessment continues to rely on patients’ subjective reports on numeric rating scales, which paradoxically also serve as training labels for machine learning models designed to detect pain without relying on self-report. Drawing on ethnography within APA research settings, mapping the field’s emergence, and historical analysis of pain measurement technologies (Tousignant 2011), the paper treats APA as an epistemic community (Knorr-Cetina 1999) organised around shared methodological problems in rendering pain computable and negotiating what counts as valid datafication.

Across these sites, the temporal unruliness of chronic pain emerges as a central application challenge for APA systems. Its fluctuating trajectories and shifting behavioural expressions disrupt the experimentally bounded temporal conditions required for algorithmic recognition. Attending to these tensions, the paper responds to this panel’s invitation by examining APA as a project of data work in the meantime, sustained through provisional proxies and experimental adjustments that keep the possibility of computable pain open (Ruckenstein and Trifuljesko 2022).