Predicting decompression surgery by applying multimodal deep learning to patients' structured and unstructured health...

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Spinal surgery prediction

Background:

• Low back pain (LBP) encompasses diverse subtypes, including lumbar disc herniation (LDH) and lumbar spinal stenosis (LSS).
• Non-surgical treatments are typically the first line of care for LDH/LSS, with surgery considered if conservative measures fail.
• Predicting surgical outcomes for LBP is complex, influenced by patient-specific health characteristics.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) model for predicting decompression surgery in patients with LDH/LSS.
• To assess the DL model's predictive performance for both early (within 2 months) and late (within 12 months) surgical interventions.
• To compare the DL model's efficacy against a benchmark LASSO logistic regression model.

Main Methods:

• Utilized large prospective datasets from four healthcare systems (8387 and 8620 patients).
• Developed a DL model incorporating patient demographics, diagnosis/procedure codes, drug names, and imaging reports.
• Evaluated model performance using classical (80% training, 20% testing) and generalizability (healthcare system split) metrics, with Area Under the Curve (AUC) as the primary measure.

Main Results:

• The DL model significantly outperformed the benchmark for early surgery prediction (AUC 0.725 vs. 0.597 classically, and superior generalizability).
• For late surgery, the DL model showed a marginal improvement classically (AUC 0.655 vs. 0.635) but was outperformed by the benchmark in generalizability.
• Performance varied between early and late surgery prediction tasks, highlighting the nuanced application of DL models.

Conclusions:

• Deep learning models show promise for predicting early surgical intervention in LBP patients.
• For late surgery prediction, DL models offer comparable performance to conventional methods, necessitating careful consideration.
• The computational expense, time investment, and interpretability challenges of DL models require thorough assessment to justify their clinical utility.