Preliminary KNN-LCB-Based Individualized Recommendation of Ultrasound Therapy Intensity for Low-Back Pain.

INTRODUCTION

Treatment‑intensity selection for conventional low‑back‑pain ultrasound therapy largely depends on empirical clinical judgment, which may cause inconsistent therapeutic outputs. This preliminary study proposes a technical scheme for intelligent recommendation of ultrasound treatment intensity using patients' baseline physiological indicators, including pain visual analog scale (VAS) score, skin‑impedance‑signal distribution features, and muscle elastic modulus.

METHODS

Thirty‑three outpatients with low back pain were enrolled from the rehabilitation outpatient department of Putuo District Central Hospital in Shanghai from October 2025 to January 2026. Baseline demographic information, pain VAS score, skin impedance at painful sites, and elastic modulus of lumbar multifidus muscle were collected before intervention. According to clinical experience, participants received ultrasound intervention at one of three intensities (3 W/cm2, 5 W/cm2, 7 W/cm2) with a fixed treatment duration of 10 minutes. Post‑treatment VAS score, skin‑impedance value and multifidus‑muscle elastic modulus at identical locations were measured. The K‑nearest neighbor (KNN) model was optimized by the lower confidence bound (LCB) strategy to construct the KNN‑LCB algorithm. Model recommendation accuracy was evaluated against clinically assigned ultrasound intensity labels.

RESULTS

After optimization by the LCB strategy, the KNN‑LCB algorithm raised the treatment‑intensity recommendation accuracy from 48.48% to 78.79%.

DISCUSSION: This preliminary exploration demonstrates that the KNN‑LCB‑based individualized recommendation approach combining clinical indices and skin‑impedance characteristics can boost the consistency and accuracy of intensity suggestion and reduce manual subjective bias. Nevertheless, this work only completes internal model validation. Clinical efficacy and safety of model‑guided ultrasound therapy have not yet been verified. Further larger‑sample external validation is required before its real‑world application for precision rehabilitation.