TOWARDS EARLY PERSONALIZED TREATMENT OF LUMBAR DISC HERNIATION: DEVELOPMENT OF A SUPERVISED DEEP LEARNING PREDICTIVE MODEL

BACKGROUND Aside from severe neurological deficits, choosing surgery versus conservative care for lumbar disc herniation (LDH) is poorly supported by current evidence. Randomized trials exclude early-surgery patients, require 6–8 weeks of conservative therapy before enrollment, and suffer high dropout/crossover rates, limiting applicability. Moreover, “conservative treatment” lacks standardization and often conflates acute and chronic presentations. As a result, guidelines defer surgery until 6 weeks–4 months of failed conservative care—potentially extending work disability, raising costs, and heightening pain chronification. PURPOSE To develop and validate a supervised deep learning model that leverages easily obtainable anamnesis data to predict the 6-month clinical outcomes of conservative treatment in patients with lumbar disc herniation. The ultimate goal is to support early clinical decision-making by identifying, at the first evaluation, patients most likely to benefit from conservative management. METHODS Patients with lumbar disc herniation are being prospectively enrolled at the Hospital of Trento, with detailed anamnesis and clinical data collected at baseline. Six months after either conservative treatment or surgery, patients are re-evaluated to determine the Oswestry Disability Index (ODI), which serves as the primary outcome. A supervised deep learning model is constructed using a compact neural network architecture, trained using stratified tenfold cross-validation with cyclical learning rates. Model performance is evaluated both as a continuous ODI prediction and within clinically meaningful score ranges. RESULTS Although data collection is ongoing, the study is designed to evaluate model performance both as a continuous ODI prediction and as a categorical prediction within 5–6 clinically meaningful score ranges. This classification approach aims to enhance interpretability and better reflect real-world decision-making thresholds. Additionally, the model will estimate the expected ODI for each patient under the alternative treatment option, allowing direct comparison between predicted outcomes of surgical and conservative care. CONCLUSIONS Our ongoing study suggests that supervised deep learning can help predict the most appropriate initial therapy for patients with lumbar disc herniation, potentially avoiding unnecessary suffering and delays. Despite a small dataset, promising results are achievable thanks to high-quality data collection, tailored neural network architectures and the efficiency of supervised training. This approach may be extended to other medical fields where decision-making is complex and large datasets are not available.