Filling the Gaps in Health Data: Using a Machine Learning Approach to Augment Partially Observed Variables Such as...

Area of Science:

• Health Informatics
• Biostatistics
• Epidemiology

Background:

• Real-world claims data often have missing behavioral confounders like smoking status.
• A specific pattern, 'missing with truncation,' occurs when 'yes' is partially observed, but 'no' is entirely missing.
• Naively treating missing smoking data as 'no' can cause significant misclassification.

Purpose of the Study:

• To evaluate transfer learning for imputing truncated smoking data in health claims.
• To compare transfer learning against a naive approach of treating missing data as absence of risk.
• To assess imputation accuracy under varying proportions of observed smokers.

Main Methods:

• A case study utilized data from the NOVELTY study (NCT02760329) with 9733 patients.
• An imputation model was trained on one data subset and evaluated on another.
• The model's performance was compared using transfer learning versus a naive 'missing equals no' approach, varying the percentage of smokers retained (q).

Main Results:

• Transfer learning achieved higher accuracy (0.89) compared to the naive approach (0.79) for imputing smoking status.
• When 90% of smokers were retained (q=90%), transfer learning's accuracy reached 0.94 versus 0.89 for the naive method.
• Transfer learning demonstrated superior accuracy when less than 80% of true smokers were recorded.

Conclusions:

• Transfer learning offers significant added value for imputing smoking data, especially when few true ever-smokers are recorded.
• The benefit of transfer learning depends on the true prevalence of smoking and the predictive model's accuracy.
• This method enhances the handling of missing behavioral confounders in large health datasets.