JMIR Publications

ABSTRACT

Background:

Non-imaging sensors combined with machine learning offer new opportunities for the digital assessment of knee osteoarthritis (KOA). However, their performance across different clinical tasks and validation conditions requires systematic evaluation.

Objective:

To evaluate the discriminative performance, validation reliability, and clinical applicability of non-imaging sensor-based machine learning for KOA identification, structural severity grading, and functional classification.

Methods:

PubMed, Web of Science, Embase, and IEEE Xplore were searched through September 16, 2026, followed by supplementary searches and citation tracking on October 2, 2026. Studies were synthesized descriptively according to clinical task, sensor modality, data partitioning strategy, and unit of prediction, while accounting for dependencies among related reports and test samples. Two reviewers used the QUADAS-3 framework to assess selected accuracy results from 82 diagnostic or grading reports, while three additional reports underwent task-adapted design assessments. Sensitivity and specificity were separately pooled in exploratory meta-analyses of pressure-sensor studies. An adapted GRADE approach was used to assess the certainty of evidence for extrapolating findings from the pressure-sensor subset to populations with clinically suspected KOA.

Results:

A total of 85 reports representing 69 related report families were included. Two representative acoustic- and inertial-sensor studies using participant-level data partitioning reported areas under the receiver operating characteristic curve (AUCs) of 0.852 and 0.910, respectively. One patient-level study of five-class Kellgren–Lawrence (KL) grading reported an accuracy of 87.0% and a macro-averaged AUC of 0.920. In the expanded pressure-sensor scenario, eight reports involving 262 participants (using the mean-feature branch of the Zhang study) yielded a pooled sensitivity of 87.6% (95% confidence interval [CI], 80.4%–94.2%) and a pooled specificity of 89.5% (95% CI, 70.3%–98.0%). In the restricted scenario involving four reports and 96 participants, pooled specificity was 64.1% (95% CI, 43.1%–86.4%). Among the 82 diagnostic or grading reports, 81 triggered at least one report-level higher-risk criterion. All 12 GRADE assessments of clinical extrapolation across the pressure-sensor scenarios and alternative analysis branches indicated very low certainty of evidence.

Conclusions:

Non-imaging sensor-based machine learning demonstrates discriminative potential for specific KOA-related tasks. However, pooled estimates from pressure-sensor studies varied according to the studies included, and their extrapolation to clinical populations remains limited. Future research should focus on clearly defined clinical tasks, prespecified data acquisition and modeling procedures, independent patient-level external validation, and evaluation of incremental value beyond existing clinical assessments. Clinical Trial: PROSPERO CRD420261506512.