Page 86 - Read Online

Page 472 Wang et al. Art Int Surg. 2025;5:465-75 https://dx.doi.org/10.20517/ais.2025.03

Table 1. List of machine learning algorithms
Challenges in clinical
Algorithm Application Strength Limitation
integration
Kalman Filter Continuous Smooth, accurate finger Requires detailed system dynamics Requires precise modeling of
Control movement predictions; real-time model; may struggle with non-linear individual dynamics; may lose
control or highly dynamic changes accuracy over time due to signal
instability
Wiener Filter Continuous Effective for continuous movement Longer history of neural data Needs frequent recalibration to
control reconstruction; statistical learning required handle signal drift
from input-output pairs
Naïve Bayes Pose High accuracy for small datasets; Assumes conditional independence, Limited by oversimplified
(NB) identification simple and computationally which may not hold in complex assumptions; struggles with noisy
efficient systems or overlapping signal patterns
Hidden Markov Pose Captures temporal dependencies Computationally intensive; requires High computational demands may
Models (HMM) identification in sequential data substantial training data for limit real-time use; sensitive to
accurate modeling variations in signal quality
HMM-NB Pose Combines spatial and temporal Complexity increases with added Complexity of integration with
identification analysis for robust predictions hidden states and larger datasets prosthetics; requires large datasets
for training robust models

to individual user patterns over time and can improve the integration of sensory feedback in prosthetic
systems. Although the current literature is largely composed of early-stage studies with small sample sizes,
future investigations should aim to improve generalizability by including larger, more diverse patient
cohorts. Stratified analyses or multicenter studies could help clarify how these factors impact the
effectiveness and reliability of RPNI-based control systems. Beyond current applications of traditional
classifiers and filters, future directions in this field will likely be driven by advanced artificial intelligence
approaches such as deep learning and reinforcement learning. Deep learning, particularly convolutional and
recurrent neural networks, offers the capacity to extract complex, high-dimensional features from EMG and
RPNI signals that may be inaccessible to conventional algorithms. Reinforcement learning, in contrast,
provides a framework for adaptive, closed-loop training in which prosthetic systems continuously refine
their performance through interaction with the user and environment. Together, these approaches could
enable prosthetic devices that not only decode user intent with higher fidelity but also adapt dynamically to
changes in signal quality, user behavior, and task context. Incorporating these techniques into future clinical
trials will be critical for achieving more robust, intuitive, and scalable prosthetic control systems. Another
critical barrier to translation is the current lack of standardized outcome measures for evaluating prosthetic
control strategies. Across published studies, metrics range from offline decoding accuracy to task-specific
functional tests (e.g., Box and Block Test, Clothespin Relocation, or custom tasks such as the Coffee Making
Task). While these provide valuable insights, the heterogeneity makes it difficult to compare results across
aggregate data for meta-analyses. Establishing a consensus on clinically meaningful endpoints, such as long-
term functional independence, user satisfaction, and quality-of-life metrics, may be essential for regulatory
approval, reimbursement, and widespread adoption. The development of standardized benchmarks and
shared datasets, as has been done in other fields of medical AI, could accelerate translation by enabling
direct comparison of algorithmic performance and prosthetic integration across diverse patient populations.

LIMITATIONS
While these studies show the promising potential of RPNIs combined with machine learning algorithms,
many remain proof-of-concept with limited sample sizes. As such, further validation in larger, more diverse
populations is essential for establishing clinical generalizability. Future work should focus on larger,
multicenter trials with more heterogeneous cohorts to enhance reproducibility and ensure the applicability
of these systems across a wider range of individuals with limb loss. Additionally, the long-term stability of