Interpretable machine-learning (in endocrinology and beyond)
0 likes67 views
1) The document discusses interpretable machine learning models, particularly for applications in endocrinology and biomedicine. It focuses on prototype-based classifiers like Learning Vector Quantization (LVQ) that are transparent and provide insight into data and problems. 2) As an example application, it analyzes steroid metabolite biomarkers from urine samples of adrenocortical tumor patients using Generalized Matrix LVQ. This identifies characteristic metabolite excretion profiles for each tumor type and relevance of individual/combined metabolites for classification. 3) The results provide domain experts diagnostic support and insight into tumor differences, demonstrating how interpretable models can analyze biomedical data and problems beyond just prediction accuracy.
![preferred: transparent/interpretable, white box
avoid blind application of ML in black box mode
interpretable machine learning
interpretable machine learning
- understand how decisions are taken
- avoid artifacts, e.g. due to hidden bias in the data
- obtain insight into the data set/problem
- posthoc simplification of the model …
accuracy is not enough [Paulo Lisboa]
… is not necessarily the goal
(e.g. basic research, biomarker identification)
😺 vs. 🐶
(sometimes it is)](https://image.slidesharecdn.com/interpretable-machine-learning-210526194528/85/Interpretable-machine-learning-in-endocrinology-and-beyond-4-320.jpg)
Interpretable machine-learning (in endocrinology and beyond)
- 1. Interpretable machine learning models in endocrinology and beyond Michael Biehl www.cs.rug.nl/~biehl Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence University of Groningen, The Netherlands Centre for Systems Modelling & Quantitative Biomedicine
- 2. supervised learning: regression / classification data: observations, e.g. vectors of num. values regression problems: predict quantitative property e.g. a b estimate weight , model: example data set classification tasks: assign data to a category x1 x2 bulls cows training: optimize parameters ⇧ model: linear separation “bull” else “cow” x1 girth x2 length x1
- 3. preferred: transparent/interpretable, white box avoid blind application of ML in black box mode interpretable machine learning popular keywords: explainable AI (XAI) fair, honest, trustworthy … AI
- 4. preferred: transparent/interpretable, white box avoid blind application of ML in black box mode interpretable machine learning interpretable machine learning - understand how decisions are taken - avoid artifacts, e.g. due to hidden bias in the data - obtain insight into the data set/problem - posthoc simplification of the model … accuracy is not enough [Paulo Lisboa] … is not necessarily the goal (e.g. basic research, biomarker identification) 😺 vs. 🐶 (sometimes it is)
- 5. • training: represent data by one or several prototypes per class • working: classify a query according to the label of the nearest prototype • decision boundaries according to (Euclidean) distances + + low storage needs little computational effort parameterized in feature space, intuitive and interpretable one intuitive framework: prototype systems for distance-based classification Learning Vector Quantization (LVQ) N-dim. feature space ? x1 x2
- 6. distance measures and relevance learning distance measure compares prototypes data points (squared) Euclidean distance - all features equally important ? - features of the same type/scale ? - are features independent ?
- 7. distance measure compares prototypes data points generalized measure relevance of a particular single feature contribution of a pair of features training: optimize prototypes and relevance matrix w.r.t. performance on training data ( objective function ) Generalized Matrix Relevance LVQ
- 8. application example: steroid metabolomics adrenocortical tumors (adenoma vs. carcinoma) www.ensat.org benign ACA malignant ACC features: 32 steroid metabolite excretion values (GC/MS) non-invasive measurement (24 hrs. urine) steroid # set of labelled example data aim: develop a tool / support system for differential diagnosis idea: analyse retrospective data by machine learning identify characteristic steroid prototypes and relevances
- 9. Generalized Matrix LVQ , ACC vs. ACA classification o pre-processing: log-transformation of excretion values • data split into 90% training, 10% validation set • training: determine prototypes and relevance matrix representative profiles (1 per class) parameterizes distance measure • validation: apply classifier to 10% hold-out data evaluates expected performance (error rates, ROC, … ) o repeat and average results over many random splits application example: steroid metabolomics
- 10. ROC characteristics clear improvement due to relevance learning on average over 1000 randomized splits 1-specificity sensitivity diagonal rel. Euclidean full matrix AUC 0.87 0.93 0.97 validation performance no relevances only diagonal full more than accuracy ?
- 11. prototypes: steroid excretion in ACA/ACC ACA ACC (z-score transformed) metabolite excretion above - average below above - average below insight: prototypes
- 12. … pairs of markers importance of single markers insight: relevance matrix 5-PT 5-PD THS facilitates selection of reduced panels with similar performance
- 13. ACA ACC relevances confirm – surprise – visualize 19 THS individually discriminative
- 14. relevances (8) 5⍺ THA (12) TH-Doc ??? confirm - surprise - visualize ACC ACA GMLVQ: multivariate analysis, discriminative combinations
- 15. ACA ACC relevance matrix is dominated by leading eigenvectors confirm – surprise - visualize • visualize data set and prototypes misclassifications? • inspect individual cases o uncertain cases outliers
- 16. GMLVQ: example of an interpretable classifier - class representatives in terms of orginal feature space - relevances of single features / combinations thereof - visualization & low-dimensional representation summary example application: steroid metabolomics based tumor classification et al. prospective
- 17. steroid metabolomics: on-going and future work - identify reduced panels of metabolites - monitoring of patients, detection of recurrences - other disorders relating to steroid metabolism … other biomedical applications GMLVQ and similar methods: - analysis of cytokine markers in rheumatoid arthritis - neuroimaging: FDG-PET scans in neurodegenerative disorders - gene expression for risk prediction in cancer - mRNA expression for the analysis of ribosome composition … methodological extensions: - high-dimensional data, heterogeneous data - modified distance measures, local relevances - probabilistic classification, forms of regression … outlook
- 18. IEEE Members News, March 2021 girth x2 length x1 some take stay home messages exploit domain knowledge (c) https://twitter.com/jessenleon
- 19. some links and example references www.cs.rug.nl/~biehl publications, news, links GMLVQ code: Matlab, Python, Java M. Biehl, B. Hammer, T. Villmann. Prototype-based models in machine learning Advanced Review in WIRES Cognitive Science, 7(2): 92-111, 2016 M. Biehl. Biomedical Applications of Prototype Based Classifiers and Relevance Learning In: International Conference on Algorithms for Computational Biology AlCoB 2017 Springer Lecture Notes in Computer Science 10252: 3-23, 2017 R. van Veen, V. Gurvits, R. Kogan, S. Meles, G.-J. de Vries, R. Renken et al. An application of Generalized Matrix Learning Vector Quantization in Neuroimaging Computer Methods and Programs in Biomedicine, Vol. 197: 105708, 2020 A. Moolla, J. de Boer, D. Pavlov, A. Amin et al. Accurate non-invasive diagnosis and staging of non-alcoholic fatty liver disease using the urinary steroid metabolome Alimentary Pharmacology and Therapeutics 51: 1188-1197, 2020