Predicting the time to get back to work using statistical models and machine learning approaches.

Area of Science:

• Computational statistics
• Machine learning applications
• Survival analysis

Background:

• The superiority of machine learning (ML) over classical statistical models in survival analyses, particularly with non-proportional hazards, remains unclear.
• The Inspiring Families programme provides support for families facing complex issues to facilitate return to work.

Purpose of the Study:

• To compare the model performance and predictive accuracy of traditional regression techniques against ML approaches.
• Utilize data from the Inspiring Families programme to evaluate different survival modeling strategies.

Main Methods:

• Compared proportional hazards models (Cox, Parmar-Royston) with ML methods: Survival penalized regression (Elastic Net), Survival Forest, and Survival Support Vector Machine.
• Explored predictors of time to return to work using 61 binary variables from 3161 participants.

Main Results:

• No single model demonstrated clear superiority; overall predictive power was low (concordance index 0.51-0.61).
• Machine learning approaches, specifically Random Survival Forest, showed a better fit (Harrell's Concordance index 0.71) compared to the Cox model (0.60).
• Identified key predictors: 'family issues and additional barriers', 'restriction of hours', 'available CV', 'self-employment considered', and 'education'.

Conclusions:

• Survival models, including those handling non-linearities, offer valuable insights and coefficient interpretation.
• Despite a better statistical fit, ML approaches did not yield substantially higher predictive power or accuracy.
• Further optimization of ML algorithms may be necessary to improve predictive capabilities in this context.