Does AI-assisted telerehabilitation improve motor outcomes in early Parkinson's?

Longer training did not change the motor endpoint

The question this trial set out to answer was whether a longer course of the same AI-assisted home training would produce a larger motor change in early Parkinson's disease. The primary outcome was the change in MDS-UPDRS Part III, the motor section of the Movement Disorder Society Unified Parkinson's Disease Rating Scale, measured from baseline to three months. Across the short-term (1 month), medium-term (2 months), and long-term (3 months) arms, the mean changes were −3.21 (95% CI −6.44 to 0.02), −1.32 (−5.37 to 2.73), and −1.13 (−7.15 to 4.89) points, with a between-group P value of 0.758.

That result is the trial's central finding, and it is negative. The confidence intervals overlap and are wide, so the data are compatible with modest benefit as well as with no effect, but the study did not demonstrate that extending training from one to three months improved motor scores. The authors attribute the absence of a detectable motor difference partly to the regimen's very low intensity and minimal volume, and suggest that targeted modules matched to individual motor phenotypes, such as balance and gait work for predominant gait disorder, would be needed to test that possibility.

What the home system actually delivered

Understanding why the motor result came out flat requires knowing what participants did. The intervention was a smartphone-based system using full-body motion capture with deep learning for human detection and 3D skeletal keypoint recognition. It gave real-time audiovisual instructions and automatically tracked joint kinematics and movement trajectories, generating feedback against predefined thresholds for joint angles, movement amplitude, and execution time. Feedback came in three levels: encouraging, positive, and high praise.

Each session consisted of six standardized exercises: mouth opening for orofacial bradykinesia, head rotation for axial mobility and rigidity, trunk rotation for axial control and postural flexibility, finger-to-thumb opposition for fine motor control, seated leg kicking for lower limb control, and sit-to-stand training for postural stability and functional mobility. Sessions typically ran 15 to 30 minutes, five days per week, giving 20, 40, or 60 sessions across the three arms. A session counted as valid only if at least 80% of prescribed exercises were completed, and adherence was recorded automatically as completed sessions relative to the prescribed schedule.

The system allowed implicit progression: because feedback thresholds were performance-based, participants were encouraged toward larger amplitude, faster execution, and better coordination over time without a fixed difficulty ladder. Concomitant Parkinson's care and treatment for comorbid conditions were left unchanged and were not adjusted by the training.

Five centers, three durations, and heavy attrition

The trial randomized 120 patients with Hoehn-Yahr stage ≤2.5 Parkinson's disease, aged 50 to 80, across five centers in a 1:1:1 ratio to the three durations, using a centrally generated computer sequence with block size six and sequentially numbered opaque sealed envelopes. Assessors were blinded to group assignment and were not involved in randomization or trial operations. Analyses followed a modified intention-to-treat principle with available-case analysis for missing data, consistent with the exploratory aim.

Attrition is the most important qualifier on every result here. Forty-nine of the 120 randomized patients (40.8%) withdrew before three months, leaving 71 in the final analysis: 25 in the short-term arm, 23 in the medium-term arm, and 23 in the long-term arm. Completers had significantly higher education levels than those lost to follow-up, and no other baseline differences were found between the two. The authors also report that dropout was associated with lower educational attainment and describe this as a digital-divide concern for equitable implementation of home digital therapeutics.

There was no usual-care or sham-control arm. That design choice means the trial can compare durations of the same intervention but cannot separate a specific effect of the AI software from non-specific trial effects, which the authors state explicitly.

The cognitive signal rests on a post hoc threshold

The one statistically significant between-group comparison came from a post hoc responder analysis, not from the prespecified primary or secondary analyses. Responders were defined pragmatically as a change exceeding 5% of the scale range: more than 3.25 points on MDS-UPDRS III, or at least 2 points on the Montreal Cognitive Assessment (MoCA). The authors note that a 2-point MoCA change is an exploratory threshold, not a validated minimal clinically important difference in this population.

On that definition, cognitive responder rates were 52.4% in the long-term group, 22.2% in the medium-term group, and 18.2% in the short-term group, with an overall χ² of 6.808 (df = 2, P = 0.033). In pairwise logistic models using the short-term group as reference, the long-term versus short-term odds ratio was 4.95 (95% CI 1.32 to 21.88, P = 0.023) unadjusted and 4.76 (1.16 to 23.27, P = 0.038) adjusted for baseline MoCA. The medium-term versus short-term comparison was not significant (unadjusted OR 1.29, 95% CI 0.26 to 6.34).

The continuous MoCA change tells a less dramatic story: −0.54 (95% CI −1.58 to 0.50) in the short-term group, 0.10 (−1.20 to 1.40) in the medium-term group, and 1.24 (0.08 to 2.40) in the long-term group, with a between-group P value of 0.076. An ANCOVA adjusting for baseline MoCA gave an adjusted long-term change of +1.06 points (95% CI 0.15 to 2.33) and an adjusted P of 0.244. The responder analysis and the continuous analysis answer related but different questions, and only the former crossed the significance line.

Motor responders and the missing motor-cognitive link

The same responder framework applied to motor scores did not separate the groups. Motor responder proportions were 28.6% in the long-term group, 27.8% in the medium-term group, and 13.6% in the short-term group, but the overall χ² test was not significant (χ² = 1.692, df = 2, P = 0.429). A higher raw proportion in the longer arms therefore does not constitute a demonstrated motor effect.

The trial also tested whether cognitive and motor changes moved together. Raw correlations between 3-month MoCA change and motor parameters were weak: gait speed R = 0.312 (P = 0.008), stride length R = 0.273 (P = 0.021), MDS-UPDRS III R = −0.270 (P = 0.023), and MDS-UPDRS total R = 0.283 (P = 0.016). None survived Benjamini-Hochberg false discovery rate correction (all corrected P = 0.098). Correlations with non-motor measures including MDS-UPDRS I, II, IV, HAMA, HAMD, ESS, PDSS-2, and activities of daily living were not significant.

The authors read this as providing no evidence for motor-cognitive coupling in their data, and they note the trial included no mechanistic markers such as functional neuroimaging or neurophysiological assessment. The correlation analysis is exploratory and corrected for multiple comparisons, so the uncorrected associations should not be treated as a mechanism.

Secondary outcomes stayed flat and no adverse events occurred

Across the secondary outcomes measured from baseline to three months, no between-group differences reached significance. MDS-UPDRS I changes were −2.52, −2.35, and −1.61 points (P = 0.767); MDS-UPDRS II changes were −1.92, −0.77, and −1.35 (P = 0.730); and total MDS-UPDRS changes were −6.75, −5.18, and −3.39 (P = 0.770). Gait parameters were similarly unseparated: gait speed P = 0.480, stride length P = 0.474, step height P = 0.753, step width P = 0.451, and turn duration P = 0.904.

Mood and other non-motor scales also showed no significant between-group differences: HAMA P = 0.502, HAMD P = 0.862, ESS P = 0.183, PDSS-2 P = 0.309, and ADL P = 0.836. Levodopa equivalent daily dosage (P = 0.613) and daily levodopa dosage (P = 0.425) did not differ either, which matters because a change in medication would complicate interpretation of motor scores.

No falls or adverse events were recorded throughout the trial. The authors attribute this safety profile to the simple rehabilitation tasks, the early disease stage, and the generally mild symptoms of the enrolled cohort, which also means the safety result should not be extended to more advanced disease or higher-intensity home programs.

What this trial can and cannot support

The most defensible reading is that a low-intensity, six-exercise AI telerehabilitation program was safe and feasible in early Parkinson's disease over one to three months, but that extending its duration did not produce a measurable motor advantage. The authors describe the intervention as a pragmatic, low-barrier approach that requires no expensive hardware or clinic visits, and they contrast it with earlier telerehabilitation studies that used immersive virtual reality, explicit cognitive drills, or higher-intensity aerobic components.

The cognitive finding should be read as a hypothesis, not an efficacy claim. It comes from a post hoc responder analysis with an exploratory threshold, in a trial whose primary endpoint was negative, with 40.8% attrition, no usual-care comparator, and confounding between intervention duration and training recency. The authors state that these observations do not permit causal inference and require validation in larger, prespecified multi-arm comparative trials.

For readers weighing this evidence, the practical distinction is between the two questions the trial answers differently. On motor outcomes, duration was not a lever: the 3-month arm's mean MDS-UPDRS III change of −1.13 points was not distinguishable from the 1-month arm's −3.21. On cognition, the only separation was the long-term versus short-term responder contrast, with an adjusted odds ratio of 4.76 and a confidence interval whose lower bound of 1.16 leaves little margin. The trial's own priority for future work follows from that split: test targeted, higher-intensity modules matched to motor phenotype in a prespecified design with an appropriate control, rather than simply extending a low-intensity general program.

Sources used in this answer