Methods:Using population-representative data from 1811 Swiss participants of the Survey of Health, Ageing and Retirement in Europe, we evaluated several ML techniques to create a PPP and we employed a commonly used, explainable ML method to identify the best-performing model. Reflecting different use scenarios, we trained three models: a simple model based on demographic data; a clinical model trained on data that are likely to be available in electronic health records; and a personalized model incorporating richer information on individual preferences.
Results:All three models outperformed the partners of index persons in our sample in accurately predicting whether the person would prefer cardiopulmonary resuscitation in the event of cardiorespiratory arrest. With a mean fivefold cross-validated accuracy of up to 70.6% (standard deviation ±1.3%), our models also performed on par with or better than typical estimates of surrogate predictive accuracy in the literature.
Conclusions:This study demonstrates that an ML-based PPP can improve surrogate decision-making. While highlighting technical and conceptual limitations, we hold this to be a major contribution to the efforts to improve care and fully respect patient autonomy.