Systematic data management for effective AI-driven decision support systems in robotic rehabilitation

Abstract

Robotic rehabilitation is becoming a standard in post-stroke physical rehabilitation, and these setups, often coupled with virtual exercises, collect a large and finely grained amount of data about patients' motor performance, in terms of kinematics and force interactions. Given the high resolution of data throughout the rehabilitation treatment, invaluable information is concealed, especially if oriented towards predictive systems and decision support systems. Nevertheless, a comprehensive understanding of how manipulating these datasets with machine-learning to produce such outputs is still missing. This study leverages comprehensive robotic-assisted rehabilitation data to systematically investigate clinical outcome predictions (FMA, ARAT and MI) and robot parameters suggestions based solely on kinematic and demographic data. Our method significantly outperforms conventional approaches on both tasks demonstrating the potential of systematic data handling in advancing rehabilitation practices. Moreover, under the explainable-AI policies, a focus on prediction power of variables and a clinical knowledge base of predicted outcome are provided.

Keywords: Decision support system; Deep neural networks; Kinematics; Machine learning; Post-stroke; Regression; Rehabilitation; Robotics.

Fig. 1

Illustration of the proposed pipeline for developing the Clinical Decision Support System. Data is collected either by the robot (kinematic data) or through an initial assessment (demographic and clinical data) or at discharge (clinical data) by therapists from 121 post-stroke patients. The collected data is preprocessed into structured data frames, which are then separated for training and testing the final models (the clinical outcome prediction model and the difficulty level recommendation model).

Fig. 2

Boxplots of (a) the normalized RMSE and (b) the normalized MAE for the three clinical scores. (c) Boxplots of the overall accuracy and (d) the F1-score for the robot parameter suggestions. Each boxplot compares the proposed LSTM-based network with a Random Forest (RF)-based model. Asterisks (***) indicate a p-value less than 0.001. (e) and (f) show dotplots of the permutation importances for the 10 most prominent features.

Fig. 3

Post processing of the results. Patients are grouped based on the clinical outcome prediction error: green if within the Minimal Clinical Important Difference (MCID) range (between -10% and +10% PD), red if over-predicted, and dark red if under-predicted. Top graphs show groups with respect to clinical scales (pre-assessment ($T_0$), post-assessment ($T_1$), and delta ($T_1-T_0$) on ARAT, FMA, and MI), demographics (age and latency), and data-related (SL, sequence length, i.e. the processed total number of days of therapy). Bottom graphs represent groups in a histogram of the whole population, based on clinical scale scores at $T_0$.

Fig. 4

Data collection protocol and detail of the clinical robotic rehabilitation setup involving virtual exercises performed with robotic assistance.

Fig. 5

The proposed algorithm dynamically shrinks time-series data to a desired length through iterative grouping and deletion. (a) The sequence is separated based on dates, and the popularity of the resulting groups is calculated. (b) The session is identified from the longest group, selecting the most prominent game. (c) This session is then removed from the dataset. (d) After each deletion, the algorithm dynamically recomputes the groups and their popularity, iterating the process until the sequence reaches the required length.

Fig. 6

3-D illustration of (a) the data structure, (b) the input shape of the clinical outcome prediction model, and (c) the input shape of the robot parameter recommendation model. Each color palette on all three dimensions indicates a physical quantity. The blue color palette corresponds to different Time-Points; the red color palette corresponds to features; and the green color palette corresponds to patients. (d) The proposed architecture of the clinical outcome prediction DST and (e) the proposed architecture of the difficulty recommendation DST.