Machine-learning model predicts catheter conversion after spinal injury

For 135 consecutive patients with early-stage spinal cord injury and neurogenic bladder, the question was not simply whether a catheter was being used, but how long it might remain in place. Jiqiang Xie and colleagues at the Fourth Military Medical University in Xi’an built a machine-learning model to estimate indwelling catheter time and predict indwelling catheter conversion within one year.

The researchers reviewed patients admitted between November 2017 and January 2026. The median indwelling catheter time was 72 days. During follow-up, 102 patients (75.56%) experienced indwelling catheter conversion.

The team tested five machine-learning survival models. The final system was a Random Survival Forest, a model that combines multiple decision trees to estimate how an event unfolds over time. SHAP — a technique that makes a model’s reasoning more interpretable — highlighted five predictors available early in care: the H-reflex, lower extremity motor score, urinary tract infection, time from lesion to rehabilitation facility and spinal cord independence measure bladder score.

The model divided patients into favorable, intermediate and unfavorable prognosis groups. Their three-month cumulative conversion rates were 97.30%, 50.23% and 2.22%, respectively. Its reported validation C-index, a measure of how well predictions distinguish between outcomes, was 0.8742; the time-dependent area under the receiver operating characteristic curve was 0.9352 at three months, 0.8834 at six months and 0.8803 at 12 months.

So what changes in practice? If independent external validation succeeds, clinicians could use a small set of early clinical measures to place patients on different bladder-management pathways instead of treating every case alike. That could help teams anticipate catheter duration and focus attention where the model signals a less favorable course. For now, the numbers come from this patient cohort: the authors describe external validation as the next condition before the tool supports personalized care pathways.