Back to 2025 Abstracts
Machine Learning Approach to Predict Pain Outcomes Following Primary Targeted Muscle Reinnervation in Amputees
Floris V. Raasveld, MD
1; Zihe Zhang, PhD
2; Benjamin R Johnston, MD, PhD
3; Anna Luan, MD
4; Arya Rao, BA
5; William Renthal, MD, PhD
6; Ian L Valerio, MD, MS, MBA
7; Kyle R. Eberlin, MD
11Massachusetts General Hospital/Harvard Medical School, Boston, MA; 2Boston Children's Hospital, Boston, MA; 3MGH, Boston, MA; 4Massachusetts General Hospital, Boston, MA; 5Massachusetts General Hospital/ Harvard Medical School, Boston, MA; 6Brigham and Women's Hospital/Harvard Medical School, Boston, MA; 7Massachusetts General Hospital | Harvard Medical School, Boston, MA
Introduction: Neuropathic pain (NP) following extremity amputation is common and may be related to symptomatic neuroma. Targeted Muscle Reinnervation (TMR) has demonstrated efficacy in its treatment, but can also be utilized primarily, as NP prophylaxis. However, TMR's effectiveness varies among patients, and it is uncertain which patients are likely to achieve the desired NP mitigation. In prior studies, we have identified patient characteristics that seem to be associated with NP mitigation. Therefore, we aimed to construct a customized Machine Learning (ML) model incorporating these patient factors to predict patient responses to Primary and Secondary TMR, to aid in improving patient selection.
Methods: Patients undergoing Primary or Secondary TMR at a tertiary care center between 2018 and 2024 were eligible for inclusion (Follow-up: >6 months). Patients were excluded if <18 years old, if they underwent minor or bilateral amputation, or if no pain data was available. Patients were identified if they achieved sustained pain mitigation (Pain remission for Secondary, Pain prophylaxis for Primary TMR, defined as a pain score of ?3 for ?3 months until final follow-up). Data on demographic, comorbidity, and surgical factors were collected through chart review. Bayesian and nonparametric modeling techniques were utilized to build a prediction model capturing the associations between patient features and the binary outcome of pain mitigation. Prediction accuracy was calculated through a relevant vector machine (RVM) model with radial basis function kernels.
Results: A total of 77 Primary and 101 Secondary TMR patients were included (median follow-up: 2.0 years) of whom 55.8% and 63.4% achieved sustained pain mitigation, respectively. The RVM training prediction accuracy were 0.86, and 0.85, and the test prediction accuracy were 0.77 (AU-ROC score 0.83), and 0.80 (AU-ROC score 0.92), respectively,
indicating that the model was able to predict with good accuracy (Table 1). In contrast, if the model would be randomly guessing, the chance levels would be 0.56, and 0.60, respectively. Discussion: This novel, custom RVM model is able to predict whether Primary and Secondary TMR patients will achieve NP mitigation or not with good accuracy. This tool could aid in preoperative counseling and improve patient selection for TMR. Such prediction models could eventually become an integral part of data-driven clinical decision-making in managing post-amputation neuropathic pain, by providing predictions of the likelihood of TMR surgical outcomes based on personalized patient features. Further, we aim to increase the accuracy of this model by ongoing prospective research on surgical outcomes.
Back to 2025 Abstracts
