Predicting deseasonalised serum 25 hydroxy vitamin D concentrations in the D-Health Trial: An analysis using boosted regression trees.

BACKGROUND

The D-Health Trial aims to determine whether monthly high-dose vitamin D supplementation can reduce the mortality rate and prevent cancer. We did not have adequate statistical power for subgroup analyses, so could not justify the high cost of collecting blood samples at baseline. To enable future exploratory analyses stratified by baseline vitamin D status, we developed models to predict baseline serum 25 hydroxy vitamin D [25(OH)D] concentration.

METHODS

We used data and serum 25(OH)D concentrations from participants who gave a blood sample during the trial for compliance monitoring and were randomised to placebo. Data were partitioned into training (80%) and validation (20%) datasets. Deseasonalised serum 25(OH)D concentrations were dichotomised using cut-points of 50, 60 and 75 nmol/L. We fitted boosted regression tree models, based on 13 predictors, and evaluated model performance using the validation data.

RESULTS

The training and validation datasets had 1788 (10.5% <50 nmol/L, 23.1% <60 nmol, 48.8 <75 nmol/L) and 447 (11.9% <50 nmol/L, 25.7% <60 nmol/L, and 49.2% <75 nmol/L) samples, respectively. Ambient UV radiation and total intake of vitamin D were the strongest predictors of 'low' serum 25(OH)D concentration. The area under the receiver operating characteristic curves were 0.71, 0.70, and 0.66 for cut-points of <50, <60 and <75 nmol/L respectively.

CONCLUSIONS

We exploited compliance monitoring data to develop models to predict serum 25(OH)D concentration for D-Health participants at baseline. This approach may prove useful in other trial settings where there is an obstacle to exhaustive data collection.