Developing diagnostic scores for prediction of clinical outcomes uses medical knowledge regarding which variables are most important and empirical/statistical learning to find the functional form of these covariates that provides the most accurate prediction (eg, highest specificity and sensitivity). Given the variables chosen by the clinician as most relevant or available due to limited resources, the job is a purely statistical one: which model, among competitors, provides the most accurate prediction of clinical outcomes, where accuracy is relative to some loss function. An optimal algorithm for choosing a model follows: (1) provides a flexible, sequence of models, which can 'twist and bend' to fit the data and (2) use of a validation procedure that optimally balances bias/variance by choosing models of the right size (complexity). We propose a solution to creating diagnostic scores that, given the available variables, will appropriately trade-off model complexity with variability of estimation; the algorithm uses a combination of machine learning, logistic regression (POLYCLASS) and cross-validation. For example, we apply the procedure to data collected from stroke victims in a rural clinic in India, where the outcome of interest is death within 30 days. A quick and accurate diagnosis of stroke is important for immediate resuscitation. Equally important is giving patients and their families an indication of the prognosis. Accurate predictions of clinical outcomes made soon after the onset of stroke can also help choose appropriate supporting treatment decisions. Severity scores have been created in developed nations (for instance, Guy's Prognostic Score, Canadian Neurological Score, and the National Institute of Health Stroke Scale). However, we propose a method for developing scores appropriate to local settings in possibly very different medical circumstances. Specifically, we used a freely available and easy to use exploratory regression technique (POLYCLASS) to predict 30-day mortality following stroke in a rural Indian population and compared the accuracy of the technique with these existing stroke scales, resulting in more accurate prediction than the existing scores (POLYCLASS sensitivity and specificity of 90% and 76%, respectively). This method can easily be extrapolated to different clinical settings and for different disease outcomes. In addition, the software and algorithms used are open-source (free) and we provide the code in the appendix.
Keywords: accuracy; mortality; prediction; prognostic model; stroke.