PACT: A predictive algorithm for automated upper-limb prosthetic socket design

Abstract

Prosthetic socket design significantly affects prosthesis comfort and abandonment [1]. However, current processes

are subjective and time-intensive, relying on extensive clinician–patient interactions that can result in long lead

times and reduced access to care. Advances in digital scanning, additive manufacturing, and machine learning

present an opportunity to streamline the initial stages of this design process, freeing up clinicians’ time to focus

on where their expertise is most impactful [2]. Despite these advances, there remains limited understanding of

digital design applications for upper-limb prostheses [3]. To address this gap, we developed and validated the

Predictive Algorithm for Customized Transradial socket design (PACT), which accurately approximates prosthe-

tist-fabricated sockets (PFSs) for individualized prosthetic care.

Nineteen (n=19) participants with transradial limb absence were recruited during regular fitting appointments.

Their limbs and corresponding PFSs were scanned using the Spectra scanner (Vorum, Canada) to create a limb–

socket library. Our algorithm takes a new client’s limb as the input, scaling and comparing it to all limbs in the

library using the average Euclidean distance to determine the best match, whose corresponding socket is used as

a basis for design. Critical areas affecting socket functionality—such as the distal end and supracondylar suspen-

sion—are then adjusted for optimal fit.

Using leave-one-out validation, sockets were predicted for all 19 limbs and compared to their respective PFSs.

Strong Spearman’s rank correlation (ρ = 0.81, P < 0.05) between scaled limb shapes and socket fits validated the

accuracy of our scaling method. The algorithm consistently predicted reasonable approximations of PFSs, with a

mean deviation of only 2.1 ± 0.6 mm across the dataset—within the acceptable range of typical inter-prosthetist

variability observed during rectification for lower-limb [4] and transradial sockets [5]. Data clustering using ma-

chine learning (DBSCAN) identified localized regions on the socket needing improvement, specifically in ante-

rior–posterior compression and the trimlines. Preliminary client fitting sessions on a subsample of participants

showed that the 3D-printed predicted sockets provided satisfactory fit and suspension.

The results demonstrate that PACT predicts reasonable approximations of PFSs, marking the first integration of

clinician-created designs into predictive modeling for prosthetic sockets. Unlike prior research using genetic al-

gorithms [6] and the eigenvector algorithm [7], PACT offers a more tailored fit by incorporating clinician-in-

formed adjustments. The widespread deployment of this algorithm could significantly reduce the burden of pros-

thesis provision on clinicians and patients. Limitations include a relatively small sample size; future iterations of

this algorithm will incorporate machine-learning techniques for limb classification and additional fitting parame-

ters to advance toward end-to-end modeling.