Interpretable machine learning risk models and rule extraction
1. Rule extraction
It is difficult to explain how prediction models work, especially when the models are based on machine learning algorithms (1). One effort to gain insight in the operation of prediction models is the use of rule extraction. In this work, we apply Orthogonal Search-Based Rule Extraction (OSRE) to extract rules from logistic regression and neural network models that estimate the risk that an ovarian tumor is malignant (2). This is an interesting method to gain insight, however the resulting rules do not yet seem practical enough for clinical use.
2. Interval-coded scoring system
In clinical practice, score systems are popular thanks to their ease of use and interpretability. However, the methods used to derive these systems are not based on statistical grounds: the predictors are divided into ad hoc intervals, and the number of intervals is often chosen without theoretical grounds. In this work, a methodological strategy was proposed to automatically obtain score systems using optimization techniques to identify the necessary cut points and intervals (3). Illustrative examples are presented for the prediction of ovarian tumor malignancy and the viability of ongoing pregnancies.
3. White-box RBF support vector machines (SVMs)
Support vector machines in combination with the Radial Basis Function (RBF) kernel are popular tools to build risk prediction models in research thanks to their flexibility and generality. The uptake of such models in clinical practice however is hampered due to their black-box nature: the user cannot defined the effect of each predictor in the prediction and hence credibility in the models predictions in difficult to obtain. To tackle this issue, this work proposes the use of a truncated RBF kernel. The latter kernel only takes the main and the two-way interaction terms of the RBF kernel into account, such that each effect used in the risk prediction can be plotted. The method was illustrated on artificial as well as bench marking data in (4) and applied to real life data to predict pregnancy viability in (5).
Van Belle V, Lisboa P. Research directions in interpretable machine learning models. Proceedings of the Eur Symp Artif Neural Netw (ESANN). 2013:533-41.
Aung MSH, Lisboa PJG, Etchells TA, Testa AC, Calster BV, Huffel SV, et al. Comparing Analytical Decision Support Models Through Boolean Rule Extraction: A Case Study of Ovarian Tumour Malignancy. Lecture Notes Comput Sci. 2007;4492:1177-86.
Van Belle VM, Van Calster B, Timmerman D, Bourne T, Bottomley C, Valentin L, et al. A mathematical model for interpretable clinical decision support with applications in gynecology. PLoS One. 2012;7(3):e34312.
Van Belle V, Lisboa P. White box radial basis function classifiers with component selection for clinical prediction models. Artif Intell Med. 2014;60(1):53-64.
Van Belle V, Lisboa P. Automated selection of interaction effects in sparse kernel methods to predict pregnancy viability. Proceedings of the IEEE Symp Comput Intell Data Mining (CIDM). 2013;26-31.