The use of machine learning for the prediction of response to follow-up in spine registries

Background: One of the main challenges in the maintenance of registries is to keep a high follow-up rate and a reliable strategy to limit dropout is currently lacking. Aim of this study was to utilize machine learning (ML) models to highlight the characteristics of patients who are most likely to drop out, and to evaluate the potential cost effectiveness of the implementation of a follow-up system based on the obtained data. Methods: All patients recruited in the local spine surgery registry were included and demographic, peri- and postoperative data were collected. Five ML models were trained and evaluated for response to follow-up prediction. Explainable and Cautious AI were then implemented to increase the trustworthiness of the model. The efficacy and cost effectiveness of the current follow-up strategy (call everybody) were compared to a strategy based on the implemented model (call only patients with high dropout risk). Results: Records from 4652 patients were available. The random forest (RF) outperformed all models in the prediction of response to follow-up. Among the considered variables, the ones that had the most weight were length of follow up, level of the main pathology and extent of surgery, SF-36 and BMI. Interpretable Decision Trees (IDT) and selective prediction models further increased the performance of the model. The cost reduction calculation predicted that implementing the developed ML model in the clinical practice would, over time, result in a reduction of costs by 31%, with only 2‰ missed calls. Conclusion: ML models can effectively identify patients with high risk of dropout. The RF model outperformed all evaluated models, and was further improved with the use of Controllable AI. The application of ML to the follow-up strategy could reduce costs and limit missed responses.