Consistency of variety of machine learning and statistical models in predicting clinical risks of individual patients: longitudinal cohort study using cardiovascular disease as exemplar.

OBJECTIVE

To assess the consistency of machine learning and statistical techniques in predicting individual level and population level risks of cardiovascular disease and the effects of censoring on risk predictions.

DESIGN

Longitudinal cohort study from 1 January 1998 to 31 December 2018.

SETTING AND PARTICIPANTS

3.6 million patients from the Clinical Practice Research Datalink registered at 391 general practices in England with linked hospital admission and mortality records.

MAIN OUTCOME MEASURES

Model performance including discrimination, calibration, and consistency of individual risk prediction for the same patients among models with comparable model performance. 19 different prediction techniques were applied, including 12 families of machine learning models (grid searched for best models), three Cox proportional hazards models (local fitted, QRISK3, and Framingham), three parametric survival models, and one logistic model.

RESULTS

The various models had similar population level performance (C statistics of about 0.87 and similar calibration). However, the predictions for individual risks of cardiovascular disease varied widely between and within different types of machine learning and statistical models, especially in patients with higher risks. A patient with a risk of 9.5-10.5% predicted by QRISK3 had a risk of 2.9-9.2% in a random forest and 2.4-7.2% in a neural network. The differences in predicted risks between QRISK3 and a neural network ranged between -23.2% and 0.1% (95% range). Models that ignored censoring (that is, assumed censored patients to be event free) substantially underestimated risk of cardiovascular disease. Of the 223 815 patients with a cardiovascular disease risk above 7.5% with QRISK3, 57.8% would be reclassified below 7.5% when using another model.

CONCLUSIONS

A variety of models predicted risks for the same patients very differently despite similar model performances. The logistic models and commonly used machine learning models should not be directly applied to the prediction of long term risks without considering censoring. Survival models that consider censoring and that are explainable, such as QRISK3, are preferable. The level of consistency within and between models should be routinely assessed before they are used for clinical decision making.