Search

Abstract The most limiting factor in heart transplantation is the lack of donor organs. With enhanced prediction of outcome, it may be possible to increase the life-years from the organs that become available. Applications of machine learning to tabular data, typical of clinical decision support, pose the practical question of interpretation, which has technical and potential ethical implications. In particular, there is an issue of principle about the predictability of complex data and whether this is inherent in the data or strongly dependent on the choice of machine learning model, leading to the so-called accuracy-interpretability trade-off. We model 1-year mortality in heart transplantation data with a self-explaining neural network, which is benchmarked against a deep learning model on the same development data, in an external validation study with two data sets: (1) UNOS transplants in 2017–2018 (n = 4750) for which the self-explaining and deep learning models are comparable in their AUROC 0.628 [0.602,0.654] cf. 0.635 [0.609,0.662] and (2) Scandinavian transplants during 1997–2018 (n = 2293), showing good calibration with AUROCs of 0.626 [0.588,0.665] and 0.634 [0.570, 0.698], respectively, with and without missing data (n = 982). This shows that for tabular data, predictive models can be transparent and capture important nonlinearities, retaining full predictive performance.

Abstract Breast cancer is the most commonly diagnosed female malignancy globally, with better survival rates if diagnosed early. Mammography is the gold standard in screening programmes for breast cancer, but despite technological advances, high error rates are still reported. Machine learning techniques, and in particular deep learning (DL), have been successfully used for breast cancer detection and classification. However, the added complexity that makes DL models so successful reduces their ability to explain which features are relevant to the model, or whether the model is biased. The main aim of this study is to propose a novel visualisation to help characterise breast cancer patients using Fisher Information Networks on features extracted from mammograms using a DL model. In the proposed visualisation, patients are mapped out according to their similarities and can be used to study new patients as a ‘patient-like-me’ approach. When applied to the CBIS-DDSM dataset, it was shown that it is a competitive methodology that can (i) facilitate the analysis and decision-making process in breast cancer diagnosis with the assistance of the FIN visualisations and ‘patient-like-me’ analysis, and (ii) help improve diagnostic accuracy and reduce overdiagnosis by identifying the most likely diagnosis based on clinical similarities with neighbouring patients.

A lack of transparency in machine learning models can limit their application. We show that analysis of variance (ANOVA) methods extract interpretable predictive models from them. This is possible because ANOVA decompositions represent multivariate functions as sums of functions of fewer variables. Retaining the terms in the ANOVA summation involving functions of only one or two variables provides an efficient method to open black box classifiers. The proposed method builds generalised additive models (GAMs) by application of L1 regularised logistic regression to the component terms retained from the ANOVA decomposition of the logit function. The resulting GAMs are derived using two alternative measures, Dirac and Lebesgue. Both measures produce functions that are smooth and consistent. The term partial responses in structured models (PRiSM) describes the family of models that are derived from black box classifiers by application of ANOVA decompositions. We demonstrate their interpretability and performance for the multilayer perceptron, support vector machines and gradient-boosting machines applied to synthetic data and several real-world data sets, namely Pima Diabetes, German Credit Card, and Statlog Shuttle from the UCI repository. The GAMs are shown to be compliant with the basic principles of a formal framework for interpretability.

In microarray studies, the number of samples is relatively small compared to the number of genes per sample. An important aspect of microarray studies is the prediction of patient survival based on their gene expression profile. This naturally calls for the use of a dimension reduction procedure together with the survival prediction model. In this study, a new method based on combining wavelet approximation coefficients and Cox regression was presented. The proposed method was compared with supervised principal component and supervised partial least squares methods. The different fitted Cox models based on supervised wavelet approximation coefficients, the top number of supervised principal components, and partial least squares components were applied to the data. The results showed that the prediction performance of the Cox model based on supervised wavelet feature extraction was superior to the supervised principal components and partial least squares components. The results suggested the possibility of developing new tools based on wavelets for the dimensionally reduction of microarray data sets in the context of survival analysis.