The CMDO Network's Health Research and AI symposium | December 13th, 2023
Cardiometabolic Health, Diabetes and obesity Research Network
Published online: December 13th, 2023Channel-Based Test-time Adaptation for Detection of Steatosis on Ultrasound B-mode Images
2Mila - Quebec AI Institute, Montreal, QC, Canada
3Flatiron Institute, New York, NY, United States
4Concordia University, Montreal, QC, Canada
Corresponding Author: Pedro Vianna, email: pedro.vianna@umontreal.ca
Abstract
Background: Deep neural networks have diverse biomedical applications, but their performance can be affected by changes in the data (different machines, populations) between training and test datasets. Test-time adaptation aims to adjust models to new data during inference. However, it may struggle with label distribution shift. Objectives: This research focuses on improving performance of models for detection of hepatic steatosis in patients with metabolic dysfunction-associated fatty liver disease. The aim is to address domain shifts using a channel selection method. Methods: We use a dataset collected in Canada as the source data, and another collected in Poland as the target data for test-time adaptation. Datasets differ by population, inclusion criteria, ultrasound scanner and settings. Based on the Wasserstein distance between adapted and unadapted channels in batch normalization layers, we propose a novel method to handle domain shift. Results: We compare our proposed method (Hybrid-TTN) with two baselines: not using any adaptation and test-time batch normalization (TTN). The experiments indicate that Hybrid-TTN presents stable accuracy fluctuations compared to no adaptation, ranging between +2.4% to -0.7% in different distribution scenarios, unlike TTN, which exhibits significant performance drops, reaching -26.2% accuracy when used in severe label shift. When adding speckle noise, Hybrid-TTN exhibits consistent positive changes, with improvements from +5.4% to +10.6%, while TTN varies between -5.8% and +11.9%. Conclusion: The study highlights that the traditional TTN might excel in scenarios without label distribution shifts but suffers from significant performance degradation in severe shifts. Conversely, the proposed Hybrid-TTN method demonstrates better adaptability, maintaining a more consistent performance across different scenarios, providing a more stable solution for domain adaptation, improving diagnostic accuracy of fatty liver. Significance: The novel proposed method enhances hepatic steatosis detection in unseen data with unknown distribution.
Patient’s ancestry prediction from ECGs using Computer Vision models
2Montreal Heart Institute, Montreal, QC, Canada
3Faculty of Medicine, Université de Montréal, Montreal, QC, Canada
4Mila - Quebec AI Institute, Montreal, QC, Canada
Corresponding Author: Uriel Nguefack Yefou, email: uriel.nguefack.yefou@umontreal.ca
Abstract
Objective: Electrocardiograms (ECG) have been the subject of several disease prediction studies using Artificial Intelligence (AI) approaches. However, since a patient's sex or age can also be predicted by such approaches, there is a danger to the fairness of such algorithms. In this project, we aim to investigate whether AI approaches on ECG can be used to predict an individual's genetic ancestry on two independent datasets. Method: To label our patients, global ancestry is inferred from a local ancestry method, RFMix, using the 1000G project reference dataset and genotyping data from the UK Biobank and MHI Biobank cohorts. Experiments were carried out on balanced datasets to perform a binary classification task: classifying the patient's ECG as European or African. An optimization method was used to determine the best model architecture and hyperparameters. The results obtained with each database were compared with a random classification to evaluate the model performance. Result: With 1112 patients in the MHI Biobank, the best model obtained after optimization was EfficientNetV2B0, with optimizer RMSprop, Binary Crossentropy as loss function, and a learning rate of 1.85e-6. Using 222 patients to test the model, we obtained 52.5±0.1% balanced accuracy against 49.5±1.6% obtained with the random method. For the UK Biobank, with 638 patients, we obtained Adam as an optimizer and a learning rate of 4.16e-6. We got 65.9±0.4% balanced accuracy on the test data (20% of the data), compared to 50.9±3% obtained with random classification. Lastly, we tried merging the two databases to increase sample size, but we found that the model was able to distinguish the origin of the ECG, which highlights a generalization issue of AI model predictions with ECG data. Conclusion: Our experiments suggest a weak but significant ancestry signal in ECGs, despite small datasets. This calls for careful examination of ethnicity biases in ECG disease prediction tasks. Furthermore, our results highlight the need to correct for the collection center for AI models to generalize on ECG data.
Interpretability of machine learning models to establish metabolomic profiles in the context of heart failure.
2Institut de Cardiologie de Montréal, Montreal, QC, Canada
3Université de Montréal, Département de Nutrition, Montreal, QC, Canada
4Université de Montréal, Département de Médecine, Montreal, QC, Canada
Corresponding Author: Cantin Baron, email: Cantin.Baron@umontreal.ca
Abstract
Heart Failure (HF) affects 64.3 million globally with a poor prognosis, highlighting the need for enhanced understanding of its biology to improve early diagnosis, patient well-being, and outcomes. In this study, we aim to derive metabolomic profiles for individuals with and without HF using interpretable machine learning (ML) models. We used targeted metabolomics data generated for 71 metabolites from individuals divided into two groups: HF patients (N=60) and controls (N=72). To predict group labels based on the metabolite levels, we used Logistic Regression (Logit), Support-Vector Machine (SVM) and Xtreme Gradient Boosting (XGB). In order to identify the most predictive metabolites for the classification task, we used two distinct strategies, a permutation-based feature importance and the Local Interpretable Model-agnostic Explanations (LIME) algorithms, to investigate group-level and individual-level interpretability, respectively. To determine feature interactions, we employed the H-Friedman statistic, aiming for a more comprehensive understanding of the network-level biology. All three classifiers predict individuals’ group labels with a good accuracy : 84.20% (σ=5.46), 85.67% (σ=6.32), 84.80.0% (σ=7.84) for Logit, SVM and XGB, respectively. In group-level interpretability analyses, all classifiers identify the same top metabolites as important features, with lignoceric acid (C24:0) being the most discriminant. The individual-level interpretability results not only corroborates the global interpretability findings, but also identifies important known biomarkers that are specific to each individual’s prediction, such as glucose and cholesterol levels. The 2-way feature interactions revealed a highly interactive network involving a dozen metabolites, highlighting crucial interactions for HF prediction. Our study shows the value of interpretable ML in metabolomics to identify potential novel biomarkers and precise biological signatures for individual health profiles.
Effectiveness of Artificial Intelligence in Cardiovascular Disease Prevention: A Systematic Review
2Faculty of Medicine, Université de Montréal, Montreal, QC, Canada
3Department of Social and Preventive Medicine, Université Laval, Quebec, QC, Canada
4Institut Cardiologie de Montréal, Montreal, QC, Canada
5Faculty of Dental Medicine, McGill University, Montreal, QC, Canada
6Mila-Quebec AI Institute, Montreal, QC, Canada
7Lady Davis Institute for Medical Research, Jewish General Hospital, Montreal, QC, Canada
Corresponding Author: Chetan Mahadev Shirvankar, email: chetan.shirvankar@mail.mcgill.ca
Abstract
Objective: Artificial Intelligence (AI) is increasingly applied in disease prevention for different reasons (e.g., optimize disease screening). AI could be useful for preventing cardiovascular diseases (CVDs). So, we evaluated the effectiveness of AI in preventing CVDs. Methods: We conducted this systematic review, searching seven databases from 1999 to October 2022 using a search strategy developed by an information specialist. Any context using AI in the prevention of CVD was considered. We included studies on the general population, patients, and healthcare professionals of any age and gender without any other qualifying criteria who receive or provide care. Two reviewers screened each paper for inclusion; a third reviewer resolved conflicts. We further conducted a subgroup analysis on studies which compare AI to standard care, such as existing risk score systems. We assessed the risk of bias using a Modified IJMEDI (International Journal of Medical Informatics) and PROBAST tool, and data was synthesized in a narrative form. We also assessed the studies for transparency, responsibility, and utility to applying AI in medicine using the Minimum Information about Clinical Artificial Intelligence Modelling (MI-CLAIM) checklist and Shifting-AI score. Results: In total, 266 articles were included from 7505 identified records. Risk prediction and early disease diagnosis were the most common prevention modalities in the articles (196/266). In conducting a detailed subgroup analysis on 18 articles with 51,687,627 participants, compared to standard care, the Framingham risk score was the most common comparator (8/18). No models were implemented in routine clinical care. AI effectiveness was not evaluated in the included studies. Conclusion: Our findings suggest that AI, specifically Machine learning models, were primarily used for risk prediction and early diagnosis of CVDs. Future research is encouraged to focus on comprehensive studies that evaluate the overall effectiveness of various AI approaches in clinical settings.
Risque d'incidence du diabète de type 2 lié à l'hématopoïèse clonale dans différents groupes ancestraux de la UK Biobank
2Département de Médecine Sociale et Préventive, Université de Montréal, Montréal, QC, Canada
3Université de Montréal, Montréal, QC, Canada
4Centre de Pharmacogénomique Beaulieu-Saucier de l’Université de Montréal/Institut de Cardiologie de Montréal, Montréal, Québec, Canada
Corresponding Author: Sam Pedro Galilee Ayivi, email: sam.pedro.galilee.ayivi@umontreal.ca
Abstract
Contexte : La présence de l'hématopoïèse clonale (HC) augmente le risque de maladies cardiovasculaires, de cancer et de mortalité. Cependant, le rôle de ces mutations dans l'incidence des maladies cardiométaboliques, en particulier au sein de populations d’origine ancestrales différentes, demeure peu étudié. Objectif: Évaluer le rôle de deux catégories de HC, les mCA (mosaic chromosome alterations) et les CHIP (clonal hematopoiesis of indeterminate potential), avec l'apparition du diabète de type 2 (DT2) au sein de populations ancestrales distinctes. Méthodes : Nous avons regroupé les individus de la UK Biobank en fonction de leur origine ancestrale déterminée génétiquement (projet pan-UK Biobank). Les données de séquençage de l'exome et de génotypage ont été utilisées pour identifier les mutations CHIP et mCA, respectivement. Les individus ayant des antécédents de cancer hématologique ou de DT2 antérieurs au recrutement ont été exclus. Nous avons évalué l'association entre CHIP ou mCA et l’incidence du DT2 à l’aide de modèles de régression de Cox ajustés. La mesure de l’importance des variables par la méthode d’impureté (Breiman et al. 1984) selon une forêt aléatoire composée d’arbres de classification CART a été appliquée aux covariables des modèles. Résultats: Cinq groupes populationnels ont été étudiés : centre et sud-asiatiques (n=7183), africains (n=5775), moyen-orientaux (n=1408), est-asiatiques (n=2533) et européens (n=422 695), avec des taux d'incidence de DT2 respectifs de 14.6%, 12%, 8.7%, 6.5% et 4.7%. Dans le sous-groupe européen, les hommes présentant un mCA sur le chromosome Y avaient un risque d'incidence de DT2 réduit de 5% (HR = 0.95, IC95% : 0.91-0.99) comparé à ceux sans mCA sur le Y. En revanche, la présence de mutations CHIP au gène DNMT3A est associé à une augmentation du risque de DT » de 10% (HR = 1.10, IC95% : 1.00-1.21). Conclusion: La compréhension du rôle et de l’impact de l’HC sur le diabète en fonction de populations diverses est essentielle pour mener des études utiles et généralisables.
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.