SEARCH · PubMed Health

Results for “machine learning explanation”

Explore indexed PubMed citations for clinical trials, systematic reviews and public health research. Read source abstracts and follow each citation to its original PubMed record.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 37 records · Page 2Linked to original sources

PubMed2026

A machine learning-based predictive model for radiosensitivity in nasopharyngeal carcinoma utilizing serum proteomics.

BACKGROUND: Nasopharyngeal carcinoma (NPC) remains highly sensitive to radiotherapy; however, radioresistance in a subset of patients leads to local recurrence and distant metastasis. Serum proteomics provides a minimally invasive approach to capturing dynamic physiological changes, and machine learning enables efficient construction of predictive models. This study aimed to develop and validate a serum proteomics–based machine-learning model for predicting radiotherapy sensitivity in nasopharyngeal carcinoma (NPC). METHODS: Pretreatment serum samples from newly diagnosed NPC patients were analyzed using SELDI-TOF-MS. Differentially expressed proteins between radiosensitive and radioresistant groups were identified using limma. GO and KEGG analyses were performed to explore functional enrichment. Twelve machine-learning algorithms were used to construct predictive models, and the top-performing models were optimized through feature selection. A Random Forest model with seven features was identified as the optimal model. External validation was performed using an independent cohort with ELISA-quantified protein levels. Model performance was assessed using Receiver operating characteristic curve (ROC), calibration analysis, decision curve analysis (DCA), and 10-fold cross-validation. SHapley Additive exPlanations (SHAP) analysis was applied for model interpretability, and the final model was deployed via a ShinyAPP. RESULTS: A total of 96 differentially expressed proteins were identified, which involved multiple function and signaling pathways. The Random Forest model demonstrated the best predictive performance, achieving an area under the curve (AUC) of 0.963 in the training set and 0.975 in the validation set. Cross-validation yielded an average AUC of 0.965. DCA indicated high clinical utility across a broad threshold range, and calibration curves showed good model agreement. Seven proteins (PLXND1, GSR, PGD, PTPRC, OR2T29, ACTG2, CHAD) were selected as final features. SHAP analysis provided global and individual-level interpretability. A web-based tool was developed to facilitate clinical application. CONCLUSION: This study establishes a robust serum proteomics–based machine-learning model capable of accurately predicting radiotherapy sensitivity in NPC. The model offers clinical interpretability and practical implementation, supporting personalized radiotherapy decision-making.

Humans↗

PubMed2026

Identification and Classification of Expressed Orphan Genes, Spurious Orphan Genes, and Conserved Genes in the Human Gut Microbiome.

Orphan genes (OGs)-genes lacking detectable homologs outside a species-are widespread in microbial genomes and are thought to contribute to their adaptation and molecular innovation. However, not all predicted OGs may represent novel functional coding sequences. False positive OGs, also called spurious OGs, can arise from gene prediction errors. We reason that OGs lacking detectable expression are more likely to be spurious. To test this, we combined large-scale metatranscriptomic profiling of the human gut microbiome with machine learning to distinguish expressed OGs from spurious ones and compare them with conserved genes (CGs) found in multiple species. Using nearly 5,000 metatranscriptome libraries, we identified ∼218,000 OGs supported by expression evidence, while ∼330,000 predicted OGs lacked detectable expression and were classified as spurious. We extracted 154 features for sequence, structural, and evolutionary properties for each gene and trained XGBoost classifiers while accounting for genomic representation. The models achieved an area under the receiver operating characteristic curve (AUC) of 0.82 in distinguishing expressed OGs from spurious OGs and an AUC of 0.93 in distinguishing expressed OGs from CGs. Interpretation based on SHAP (SHapley Additive exPlanations) revealed clear biological signals. Particularly, expressed orphans were present in more genomes than spurious ones, and expressed OGs were shorter than CGs. This work improves OG discovery and suggests that expressed OGs differ systematically from CGs and spurious OGs in sequence composition, structural constraints, and evolutionary signals.

Humans↗

PubMed2000

Visual management of large scale data mining projects.

This paper describes a unified framework for visualizing the preparations for, and results of, hundreds of machine learning experiments. These experiments were designed to improve the accuracy of enzyme functional predictions from sequence, and in many cases were successful. Our system provides graphical user interfaces for defining and exploring training datasets and various representational alternatives, for inspecting the hypotheses induced by various types of learning algorithms, for visualizing the global results, and for inspecting in detail results for specific training sets (functions) and examples (proteins). The visualization tools serve as a navigational aid through a large amount of sequence data and induced knowledge. They provided significant help in understanding both the significance and the underlying biological explanations of our successes and failures. Using these visualizations it was possible to efficiently identify weaknesses of the modular sequence representations and induction algorithms which suggest better learning strategies. The context in which our data mining visualization toolkit was developed was the problem of accurately predicting enzyme function from protein sequence data. Previous work demonstrated that approximately 6% of enzyme protein sequences are likely to be assigned incorrect functions on the basis of sequence similarity alone. In order to test the hypothesis that more detailed sequence analysis using machine learning techniques and modular domain representations could address many of these failures, we designed a series of more than 250 experiments using information-theoretic decision tree induction and naive Bayesian learning on local sequence domain representations of problematic enzyme function classes. In more than half of these cases, our methods were able to perfectly discriminate among various possible functions of similar sequences. We developed and tested our visualization techniques on this application.

Alcohol Dehydrogenase↗

PubMed2025

Integrating explainable AI with multiomics systems biology and EHR data mining for personalized drug repurposing in Alzheimer's disease.

Alzheimer's disease (AD) is characterized by region- and patient-specific molecular heterogeneity, which hinders therapeutic design. In this study, we introduce PRISM-ML (PRecision-medicine using Interpretable Systems and Multiomics with Machine Learning), an open-source integrated analysis pipeline that combines interpretable machine learning with systems biology and electronic health record (EHR) data mining to elucidate the molecular diversity of AD and predict promising drug repurposing opportunities. First, we integrated and harmonized transcriptomic (bulk RNA-seq) and genomic (genome-wide association study) data from 2105 brain samples, each with matched data from the same individual (1363 AD patients, 742 controls; nine tissues), sourced from three independent studies. Random forest classifiers with SHapley Additive exPlanations (SHAP) identified patient-specific biomarkers; unsupervised clustering resolved 36 molecularly distinct "subtissues" (clusters of samples); and gene-gene co-expression networks prioritized 262 high-centrality bottleneck genes as putative regulators of dysregulated pathways. Next, knowledge graph-based drug repurposing predicted six FDA-approved drugs that simultaneously target multiple bottleneck genes and multiple AD-relevant pathways. Notably, in a large U.S. de-identified insurance-claims database (n = 364733), exposure to promethazine, one of the candidate drugs, was associated with a 57-62 % lower incidence of AD versus an active antihistamine comparator (adjusted hazard ratio 0.38; inverse-probability weighted 0.43; both p < 0.001), providing real-world support for its repurposing potential. In summary, PRISM-ML, as an explainable multi-omics analysis pipeline, is readily transferable to other complex diseases, advancing precision medicine.

Computational Biology↗

PubMed2025

Integrating explainable artificial intelligence with multiomics systems biology and electronic health record data mining for personalized drug repurposing in Alzheimer's disease.

Alzheimer's disease (AD) is characterized by region- and patient-specific molecular heterogeneity, which hinders therapeutic design. In this study, we introduce PRISM-ML (PRecision-medicine using Interpretable Systems and Multiomics with Machine Learning), an open-source integrated analysis pipeline that combines interpretable machine learning with systems biology and electronic health records data mining to elucidate the molecular diversity of AD and predict promising drug repurposing opportunities. First, we integrated and harmonized transcriptomic (bulk RNA-seq) and genomic (genome-wide association study) data from 2105 brain samples, each with matched data from the same individual (1363 AD patients, 742 controls; 9 tissues), sourced from three independent studies. Random forest classifiers with SHapley Additive exPlanations identified patient-specific biomarkers; unsupervised clustering resolved 36 molecularly distinct subtissues (defined as clusters of samples within a brain tissue that share a specific expression pattern); and gene-gene coexpression networks prioritized 262 high-centrality bottleneck genes as putative regulators of dysregulated pathways. Next, knowledge graph-based drug repurposing predicted six Food and Drug Administration (FDA)-approved drugs that simultaneously target multiple bottleneck genes and multiple AD-relevant pathways. Notably, in a large US de-identified insurance-claims database (n = 364 733), exposure to promethazine, one of the candidate drugs, was associated with a 57%-62% lower incidence of AD versus an active antihistamine comparator (adjusted hazard ratio 0.38; inverse-probability weighted 0.43; both P < .001), providing real-world support for its repurposing potential. In summary, PRISM-ML, as an explainable multiomics analysis pipeline, is readily transferable to other complex diseases, advancing precision medicine.

Alzheimer Disease↗

PubMed1990

Initiating ethanol drinking in a simian social group in a naturalistic setting.

We examined in nine group-living, male Japanese Snow monkeys the initiation of alcohol-drinking behavior in an enriched environment where both social and nonsocial stimuli could influence drinking. The monkeys usually could move freely between an indoor shelter and a large outdoor corral, which contained three drinkometers. During daily 2-hr sessions in pre- and post-training periods (with food and water available ad libitum), the drinkometers held (on different days) Koolaid-saccharin, water, water with 5% ethanol, Koolaid-saccharin with 5% ethanol, or Koolaid-sucrose with acetic acid (matched in calories and palatability to Koolaid-saccharin with 5% ethanol). Between pre- and post-training periods was a long training period, in which Koolaid-saccharin with various ethanol concentrations was presented with the daily food ration in 2-hr drinking sessions. On other selected training days, water with 5% ethanol was presented. During training sessions some monkeys drank high doses; others did not. During the period of peak drinking, the daily mean ethanol consumption ranged among animals from 0.54-1.99 ml/kg. Blood ethanol concentrations then sometimes exceeded 100 mg/dl. During the post-training period, with return to ad libitum food and water, consumption declined from these peaks, but remained significantly higher than pretraining consumption; the ethanol solution was established as a reinforcer. Monkeys differed significantly in the extent of this pre/post increase. In both pre- and post-training periods, consumption was significantly greater when all three drinkometers operated, compared with only one. On pretraining days when only one drinkometer operated, more dominant animals drank significantly more than less dominant animals; this difference disappeared by the post-training period, as less dominant animals learned to use the drinkometers at times when the dominant animals eschewed the machines. Pretraining consumption did not predict (among animals) post-training consumption. Before, during, and after training, these animals drank less of both alcohol and control solutions than did members of another species, which we had studied in smaller indoor pens. We discuss possible explanations and implications.

Alcohol Drinking↗

PubMed2004

Adaptive user displays for intelligent tutoring software.

Intelligent tutoring software (ITS) holds great promise for K-12 instruction. Yet it is difficult to obtain rich information about users that can be used in realistic educational delivery settings--public school classrooms--in which eye tracking and other user sensing technologies are not suitable. We are pursuing three "cheap and cheerful" strategies to meet this challenge in the context of an ITS for high school math instruction. First, we use detailed representations of student cognitive skills, including tasks to assess individual users' proficiency with abstract reasoning, proficiency with simple math facts and computational skill, and spatial ability. Second, we are using data mining and machine learning algorithms to identify instructional sequences that have been effective with previous students, and to use these patterns to make decisions about current students. Third, we are integrating a simple focus-of-attention tracking system into the software, using inexpensive, web cameras. This coarse-grained information can be used to time the display of multimedia hints, explanations, and examples when the user is actually looking at the screen, and to diagnose causes of problem-solving errors. The ultimate goal is to create non-intrusive software that can adapt the display of instructional information in real time to the user's cognitive strengths, motivation, and attention.

Algorithms↗

PubMed1998

Population size and quality in genetics-based rule learning from medical data.

Population size and quality are parameters which control the performance of genetic algorithms. We researched these parameters in a genetic-based machine learning system Galactica which was used to discover the differential diagnostic rules for female urinary incontinence from case data. The performance of the system was measured with on-line and off-line criteria. Surprisingly, randomly generated small populations (30 and 70 rules) did not promote the best on-line performance as earlier results suggested. Probable explanation is the lack of diversity in initial populations. The seeding of population with positive learning examples was used to obtain more divergent populations. As expected, the seeding increased the on-line performance of small populations. The results are mainly in accord with the earlier results indicating that large randomly generated populations (150 rules) lead to the better off-line performance. Again, the seeding of small populations was successful producing even the better off-line performance than a large population. In conclusion, the seeding allowed the small populations to converge to good rules in relatively short period of time.

Algorithms↗

PubMed2005

Modeling clinical judgment and implicit guideline compliance in the diagnosis of melanomas using machine learning.

We explore several machine learning techniques to model clinical decision making of 6 dermatologists in the clinical task of melanoma diagnosis of 177 pigmented skin lesions (76 malignant, 101 benign). In particular we apply Support Vector Machine (SVM) classifiers to model clinician judgments, Markov Blanket and SVM feature selection to eliminate clinical features that are effectively ignored by the dermatologists, and a novel explanation technique whereby regression tree induction is run on the reduced SVM model's output to explain the physicians' implicit patterns of decision making. Our main findings include: (a) clinician judgments can be accurately predicted, (b) subtle decision making rules are revealed enabling the explanation of differences of opinion among physicians, and (c) physician judgment is non-compliant with the diagnostic guidelines that physicians self-report as guiding their decision making.

Artificial Intelligence↗

PubMed1994

Induction of rules for biological macromolecule crystallization.

X-ray crystallography is the method of choice for determining the 3-D structure of large macromolecules at a high enough resolution. The rate limiting step in structure determination is the crystallization itself. It takes anywhere between a few weeks to several years to obtain macromolecular crystals that yield good diffraction patterns. The theory of forces that promote and maintain crystal growth is preliminary, and crystallographers systematically search a large parameter space of experimental settings to grow good crystals. There is a wealth of experimental data on crystal growth most of which is in paper laboratory notebooks. Some of the data has been gathered in electronic form, e.g., the Biological Macromolecular Crystallization Database (BMCD) which is a repository of successful experimental conditions for growing over 800 different macromolecules (Gilliland 1987). Crystallographers are in need of computational tools to gather and analyze past data to design new crystal growth trails. We are building the Crystallographer's Assistant (CA) to help crystallographers record and maintain experimental context in electronic form, offer suggestions on experimental conditions that are likely to be successful, and provide explanations for failed experiments. As an initial step in this project, we have applied RL, an inductive learning program, to the BMCD. In this paper we report initial experiments and findings in applying RL to the BMCD. From the point of view of crystallography, we have discovered possibly significant new empirical relationships in crystal growth. From the point of view of machine learning, our work suggests refinements of existing methods for incorporating detailed domain knowledge into inductive analysis techniques.

Animals↗

PubMed2025

Prediction of gene expression using histone modification patterns extracted by Particle Swarm Optimization.

MOTIVATION: Histone modifications play an important role in transcription regulation. Although the general importance of some histone modifications for transcription regulation has been previously established, the relevance of others and their interaction is subject to ongoing research. By training Machine Learning models to predict a gene's expression and explaining their decision making process, we can get hints on how histone modifications affect transcription. In previous studies, trained models were either hardly explainable or the models were trained solely on the abundance of histone modifications. Based on other studies, which used histone modification patterns, rather than their abundance, to identify potential regulatory elements, we hypothesize the histone modification pattern in a gene's promoter to be more predictive for gene expression. We used an optimization algorithm to extract predictive histone modification profiles. RESULTS: Our algorithm called PatternChrome achieved an average area under curve (AUC) score of 0.9029 over 56 samples for binary classification, outperforming all previous algorithms for the same task. We explained the models decisions to deduce the effect of specific features, certain histone modifications or promoter positions on transcription regulation. Although the predictive histone modification patterns were extracted for each sample separately, they can be used to predict gene expression in other samples, implying that the created patterns are largely generalizable. Interestingly, the impact of histone modifications on gene regulation appears predominantly indifferent to cellular specificity. Through explanation of the classifier's decisions, we substantiate established literature knowledge while concurrently revealing novel insights into the intricate landscape of transcriptional regulation via histone modification. AVAILABILITY AND IMPLEMENTATION: The code for the PatternChrome algorithm, the scripts for the analyses and the required data can be found at (https://gitlab.gwdg.de/MedBioinf/generegulation/patternchrome).

Humans↗

PubMed2006

Applying hybrid reasoning to mine for associative features in biological data.

We develop the means to mine for associative features in biological data. The hybrid reasoning schema for deterministic machine learning and its implementation via logic programming is presented. The methodology of mining for correlation between features is illustrated by the prediction tasks for protein secondary structure and phylogenetic profiles. The suggested methodology leads to a clearer approach to hierarchical classification of proteins and a novel way to represent evolutionary relationships. Comparative analysis of Jasmine and other statistical and deterministic systems (including Explanation-Based Learning and Inductive Logic Programming) are outlined. Advantages of using deterministic versus statistical data mining approaches for high-level exploration of correlation structure are analyzed.

Algorithms↗

PubMed2026

Machine Learning-Driven Prediction of Coronary Artery Disease Risk Based on UK Biobank Plasma Proteomics.

BACKGROUND: Coronary artery disease (CAD) is a leading global cause of mortality, yet the predictive accuracy of conventional risk models is limited. Here, we integrate conventional risk factors, polygenic risk scores, and large-scale proteomics to develop a unified model for enhanced CAD risk prediction. METHODS: Using data from UK Biobank, participants with plasma proteomics and genetic risk data were included after excluding prevalent CAD. Participants from England were split into training (n=32 330) and internal validation (n=13 857) sets, and Scotland/Wales participants formed an external validation set (n=5775). Incident CAD was ascertained from linked health records. A 202-protein proteomic risk score was derived by least absolute shrinkage and selection operator Cox regression, and CatBoost models were trained using conventional risk factors alone and with incremental addition of polygenic risk scores and protein proteomic risk scores; Shapley Additive Explanations-guided forward selection identified a compact protein panel. RESULTS: Across cohorts, the median age was 58 years and ∼45% were men. Protein proteomic risk score was dose-dependently associated with CAD risk. Compared with conventional risk factors alone, integrating polygenic risk scores and protein proteomic risk scores improved discrimination, with the area under the curve increasing from 0.750 (95% CI, 0.732-0.767) to 0.789 (95% CI, 0.772-0.805) in internal validation and from 0.717 (95% CI, 0.683-0.750) to 0.762 (95% CI, 0.732-0.791) in external validation. A 9-protein panel (GDF15 [growth differentiation factor 15], MMP12 [matrix metalloproteinase 12], NPPB [natriuretic peptide B], PGF [placental growth factor], REN [renin], ADGRG2 [adhesion G-protein coupled receptor], ACE2 [angiotensin-converting enzyme 2], CDCP1 [CUB domain-containing protein 1], CXCL17 [C-X-C motif chemokine ligand 17)]) captured most proteomic predictive information. CONCLUSIONS: Our findings demonstrate that integrating conventional risk factors, polygenic risk scores, and proteomic data improves CAD risk prediction. This study highlights the utility of proteomics in precision cardiovascular medicine and simplified risk stratification tools.

Humans↗

PubMed2001

Computerized radiographic mass detection--part II: Decision support by featured database visualization and modular neural networks.

Based on the enhanced segmentation of suspicious mass areas, further development of computer-assisted mass detection may be decomposed into three distinctive machine learning tasks: 1) construction of the featured knowledge database; 2) mapping of the classified and/or unclassified data points in the database; and 3) development of an intelligent user interface. A decision support system may then be constructed as a complementary machine observer that should enhance the radiologists performance in mass detection. We adopt a mathematical feature extraction procedure to construct the featured knowledge database from all the suspicious mass sites localized by the enhanced segmentation. The optimal mapping of the data points is then obtained by learning the generalized normal mixtures and decision boundaries, where a is developed to carry out both soft and hard clustering. A visual explanation of the decision making is further invented as a decision support, based on an interactive visualization hierarchy through the probabilistic principal component projections of the knowledge database and the localized optimal displays of the retrieved raw data. A prototype system is developed and pilot tested to demonstrate the applicability of this framework to mammographic mass detection.

Artificial Intelligence↗

PubMed2026

Genomic signatures associated with epidemiologically defined high-risk pathogenic Escherichia coli isolates identified by interpretable machine learning.

Pathogenic Escherichia coli is a major cause of foodborne illness worldwide and includes strains capable of causing severe disease. To establish a genome-informed framework for foodborne outbreak surveillance, we analyzed 1,029 E. coli isolates from clinical, food, livestock, and environmental sources using whole-genome sequencing. Pathogenic isolates obtained from human clinical cases or linked to documented outbreaks were classified as epidemiologically defined high-risk (EpiHR), whereas the remaining pathogenic isolates were classified as non-EpiHR. Virulence-associated genomic features were extracted using a bioinformatics pipeline, and four machine learning (ML) algorithms, including gradient boosting machine, random forest (RF), and support vector machines with linear and radial basis function kernels, were evaluated. Among them, the RF model showed the best performance, achieving an area under the curve (AUC) of 0.98 and accuracy of 0.93 in 10-fold cross-validation. Additional leave-one-group-out validation showed retained discrimination across held-out sequence types and serotypes, although performance was reduced when isolates were grouped by isolation source. Evaluation using an independent test dataset of 1,908 publicly available pathogenic E. coli genomes showed an AUC of 0.97 and a sensitivity of 0.98. Feature importance analysis using Shapley additive explanations identified influential predictive features, including traT, etpB, and enterotoxin-associated genes. A reduced 10-feature model achieved an AUC of 0.79 in the independent test dataset, supporting its exploratory use for future simplified screening approaches. These results indicate that genome-based ML provides a sensitive framework for surveillance-oriented prioritization of EpiHR pathogenic E. coli isolates, with model predictions interpreted together with epidemiological information.

Escherichia coli↗

PubMed2000

Boosting neural networks.

Boosting is a general method for improving the performance of learning algorithms. A recently proposed boosting algorithm, AdaBoost, has been applied with great success to several benchmark machine learning problems using mainly decision trees as base classifiers. In this article we investigate whether AdaBoost also works as well with neural networks, and we discuss the advantages and drawbacks of different versions of the AdaBoost algorithm. In particular, we compare training methods based on sampling the training set and weighting the cost function. The results suggest that random resampling of the training data is not the main explanation of the success of the improvements brought by AdaBoost. This is in contrast to bagging, which directly aims at reducing variance and for which random resampling is essential to obtain the reduction in generalization error. Our system achieves about 1.4% error on a data set of on-line handwritten digits from more than 200 writers. A boosted multilayer network achieved 1.5% error on the UCI letters and 8.1% error on the UCI satellite data set, which is significantly better than boosted decision trees.

Algorithms↗

PubMed2025

Assessing individual genetic susceptibility to metabolic syndrome: interpretable machine learning method.

BACKGROUND: Genome-wide association studies have provided profound insights into the genetic aetiology of metabolic syndrome (MetS). However, there is a lack of machine-learning (ML)-based predictive models to assess individual genetic susceptibility to MetS. This study utilized single-nucleotide polymorphisms (SNPs) as variables and employed ML-based genetic risk score (GRS) models to predict the occurrence of MetS, bringing it closer to clinical application. METHODS: Feature selection was performed using Least Absolute Shrinkage and Selection Operator. Six ML algorithms were employed to construct GRS models. A fivefold cross-validation was utilized to aid in the internal validation of models. The receiver operating characteristic (ROC) curve was used to select the better-performing GRS model. The SHapley Additive exPlanations (SHAP) was then applied to interpret the model. After extracting GRS, stratified analysis of BMI, age and gender was performed. Finally, these conventional risk factors and GRS were integrated through multivariate logistic regression to establish a combined model. RESULTS: A total of 17 SNPs were selected for analysis. Among the GRS models, the extreme gradient boosting (XGBoost) model demonstrated superior discriminative performance (AUC = 0.837). The XGBoost's optimal robustness was also validated through five-fold cross-validation (mean ROC-AUC = 0.706). The XGBoost-based SHAP algorithm not only elucidated the global effects of 17 SNPs across all samples, but also described the interaction between SNPs, providing a visual representation of how SNPs impact the prediction of MetS in an individual. There was a strong correlation between GRS and MetS risk, particularly observed among young individuals, males and overweight individuals. Furthermore, the model combining conventional risk factors and GRS exhibited excellent discriminative performance (AUC = 0.962) and outstanding robustness (mean ROC-AUC = 0.959). CONCLUSION: This study established a reliable XGBoost-based GRS model and a GRS prediction platform (https://metabolicsyndromeapps.shinyapps.io/geneticriskscore/) to assess individual genetic susceptibility to MetS. This model has high interpretability and can provide personalized reference for determining the necessity of primary prevention measures for MetS. Additionally, there may be interactions between traditional risk factors and GRS, and the integration of both in a comprehensive model is useful in the prediction of MetS occurrence.

Humans↗

PubMed1999

Case-based explanation of non-case-based learning methods.

We show how to generate case-based explanations for non-case-based learning methods such as artificial neural nets or decision trees. The method uses the trained model (e.g., the neural net or the decision tree) as a distance metric to determine which cases in the training set are most similar to the case that needs to be explained. This approach is well suited to medical domains, where it is important to understand predictions made by complex machine learning models, and where training and clinical practice makes users adept at case interpretation.

Artificial Intelligence↗