Palpation Alters Auditory Pain Expressions with Gender-Specific Variations in Robopatients

TL;DR

This study presents an early attempt to use human-in-the-loop reinforcement learning to co-optimize haptic input and auditory pain expression in robopatients, and shows that pain sound perception exhibits saturation at lower forces with gender-specific thresholds.

Abstract

Diagnostic errors remain a major cause of preventable mortality, particularly in resource limited settings. Medical training simulators, including robopatients, help reduce such errors by replicating patient responses during procedures such as abdominal palpation. However, generating realistic multimodal feedback especially auditory pain expressions remains challenging due to the complex, nonlinear relationship between applied palpation forces and perceived pain sounds. The high dimensionality and perceptual variability of pain vocalizations further limit conventional modeling approaches. We propose a novel experimental paradigm for adaptive pain expressivity in robopatients that dynamically generates auditory pain responses to palpation forces using human in the loop machine learning. Specifically, we employ Proximal Policy Optimization (PPO), a reinforcement learning algorithm suited for continuous control, to iteratively refine pain sound generation based on real time human evaluative feedback. The system initializes randomized mappings between force inputs and sound outputs, and the learning agent progressively adjusts them to align with human perceptual preferences. Results show that the framework adapts to individual palpation behaviors and subjective sound preferences while capturing a broad range of perceived pain intensities, from mild discomfort to acute distress. We also observe perceptual saturation at lower force ranges, with gender specific thresholds in pain sound perception. This work demonstrates the feasibility of human in the loop reinforcement learning for co-optimizing haptic input and auditory pain expression in medical simulators, highlighting the potential of adaptive and immersive platforms to enhance palpation training and reduce diagnostic errors.

Figures (9)

• Figure 1: Multimodal pain synthesis for robopatient. The system comprises three main components: the user (ideally a physician), the robopatient, and the human patient. The human patient under examination may exhibit various physiological abnormalities, psychological conditions, and individual characteristics, all of which influence the dimensionality of pain expression in response to palpation. The robopatient serves as an adaptive intermediary, learning appropriate pain responses from the user feedback and aiming to replicate human pain expressions (including vocal pain sounds and facial pain expressions) realistically. During palpation, robot's abdominal phantom captures user's haptic input, which is recorded and processed through the robopatient’s internal palpation-pain mapping model. Then the robot outputs corresponding pain sounds and facial expressions through MorphFace lalitharatne2021morphface. The user provides feedback on the robopatient’s generated pain responses, which are iteratively refined by comparing palpation data with user feedback. This interaction loops continuously, enabling the robopatient to progressively learn and refine pain expressions while accurately capturing palpation details from the user. This concept was adapted from lalitharatne2022face and modified to fit the new paradigm.
• Figure 2: Physical robot and the experiment procedure A) Robopatient setup: A scenario where the participant is palpating on robopatient is shown. B) Data flow within the robopatient setup during the experiment. The interaction between the robot and the user starts as the user palpates on the abdominal phantom (Steps 1, 2). The force applied by the user has to reach a certain level indicated by a progress bar next to the robot's face (step 3). The progress bar is initially green and turns red when the required minimum amount of force is applied. Based on the palpation force, the robot generates random facial expressions and sounds independent to each other (step 4). Then the user listens to the pain sounds generated by the robot for the palpation force (step 5) and provides feedback on whether they agree with the robot’s mapping of pain sounds to palpation (step 6). This feedback updates the policy used by the PPO to generate future responses (step 7). This completes the one interaction cycle between the user and robot. Then robopatient learns the appropriate combinations of pitch and amplitude for pain sounds and optimizes its response to maximise reward.
• Figure 3: Pain expressions used in the experiment. Audio tracks to represent pain in A) males and B) females. These can be downloaded from https://github.com/ChapaSiri12/Robopatient_PPO where all the codes used in the study can be found. Selected frames of C) male and D) female facial pain expressions (frame 0-neutral and frame 100-intense pain) for each gender. Selected consecutive images out of 100 were played in order to create different intensities of pain expressions for each gender are shown.
• Figure 4: Process flow diagram of the PPO algorithm.
• Figure 5: A) The force distribution for a given attempt for palpation is plotted against time. This corresponds to the target force of 15N. B) Box and whisker plots indicate the spread of actual palpation forces observed for each target force for the robot's pain expressions of each gender.