| Titre : | Data Driven Cancer Prognosis using Machine Learning: Towards Clinical eXplainable AI |
| Auteurs : | Ikram Maouche, Auteur ; Sadek Labib Terrissa, Directeur de thèse |
| Type de document : | Thése doctorat |
| Editeur : | Biskra [Algérie] : Faculté des Sciences Exactes et des Sciences de la Nature et de la Vie, Université Mohamed Khider, 2026 |
| Format : | 1 vol. (281 p.) / couv. ill. en coul / 30 |
| Langues: | Français |
| Langues originales: | Anglais |
| Résumé : |
Breast cancer remains a major global health challenge and the leading cause of cancer-related mortality among women, primarily due to disease recurrence after a period of remission. The heterogeneous nature of breast cancer poses substantial challenges for accurate individualised prognosis prediction. While machine learning-based computer-aided prognosis models have shown strong potential to improve personalised prognosis estimation, their clinical adoption is hindered by multiple data and model related issues. The most well-known data-related issue is class imbalance, while model-related issues include multimodal data integration and the lack of explainability of the CAP model decisions. This thesis addresses these challenges by proposing explainable machine learning approaches for breast cancer prognosis prediction, introducing two complementary contributions.The first contribution focuses on patient-level explainable prediction of breast cancer recurrence using clinicopathological data. A cost-sensitive CatBoost (CS-CatBoost) classifier combined with genetic algorithm-based feature selection is proposed to handle class imbalance and improve recurrence prediction. To enhance clinical trust in the predictions of CS-CatBoost, patient-level explainability methods LIME and SHAP are employed to quantify the impact of patient characteristics on model predictions. The proposed approach achieved a high sensitivity for both recurrence and no-recurrence classes. In addition, the explainability methods provided stable, faithful explanations that identified clinically relevant predictive factors. Although clinicopathological information has shown to be useful for this purpose, it offers only a limited perspective on the intricate biological processes driving this disease. Individuals with similar clinicopathological profiles can display treatment responses and outcomes. In addition, the provided feature importance scores are primarily descriptive and do not provide actionable guidance that could directly assist in decision-making for prognosis improvement. The second contribution addresses the integration of heterogeneous multimodal data for breast cancer recurrence prediction. An explainable Multimodal Co-attention-based Graph Convolutional Network (M-CoGCN) is proposed to learn inter-modality information ofclinicopathological and multi-omics data. To further support clinical decision-making, a novel counterfactual explainability approach based on NSGA-II is introduced. It generates actionable counterfactual scenarios that suggest minimal changes to patient characteristics to improve prognostic outcomes. M-CoGCN achieved strong predictive performance, outperforming several state-of-the-art models. The results highlighted the dominance of clinicopathological characteristics in guiding prediction, with multi-omics providing complementary discriminative information. Moreover, the proposed counterfactual explanation approach generated clinically plausible and actionable recommendations for improved prognosis, assisting clinicians in their decision-making |
| Sommaire : |
Abstract ii ٌۘ ڲڪiii Résumé iv Acknowledgements v 1 General Introduction 1 1.1 Context and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Purpose and Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.4 Manuscript Organisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2 Computer-Aided Breast Cancer Prognosis 7 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2.2 Medical Background: Breast Cancer . . . . . . . . . . . . . . . . . . . . . . . . . 7 2.2.1 What is Breast Cancer? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.2.2 Breast Cancer Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.2.3 Breast Cancer Treatment . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.2.4 Breast Cancer Prognosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.2.4.1 Prognostic Assessment Process . . . . . . . . . . . . . . . . . . . 11 2.2.4.2 Clinical Endpoints . . . . . . . . . . . . . . . . . . . . . . . . . . 13 2.2.4.3 Prognostic Factors and Predictive Factors . . . . . . . . . . . . . 13 2.2.5 Towards Precision Prognosis . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.3 Technical Background: Computer-Aided Prognosis . . . . . . . . . . . . . . . . . 15 2.3.1 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.3.1.1 Data Cleaning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.3.1.2 Data Transformation . . . . . . . . . . . . . . . . . . . . . . . . 16 2.3.2 Data Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.3.2.1 Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.3.2.2 Dimensionality Reduction . . . . . . . . . . . . . . . . . . . . . . 20 2.3.2.3 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . 20 2.3.3 Prognostic Model Construction . . . . . . . . . . . . . . . . . . . . . . . . 20viii 2.3.3.1 Classifier Selection . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.3.3.2 Model Training and Assessment . . . . . . . . . . . . . . . . . . 28 2.4 Challenges in Data-driven Breast Cancer Prognostic Modelling . . . . . . . . . . 31 2.4.1 Data-related Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.4.1.1 Class Imbalance . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.4.1.2 Data Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.4.2 Model-related Challenges and Issues . . . . . . . . . . . . . . . . . . . . . 33 2.4.2.1 Bias-Variance Trade-off . . . . . . . . . . . . . . . . . . . . . . . 33 2.4.2.2 Accuracy-Explainability Trade-off . . . . . . . . . . . . . . . . . 34 2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 3 Explainable AI for Medical Prognosis 36 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 3.2 Theoretical Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 3.3 The Need for Explainable AI in Medical Prognosis . . . . . . . . . . . . . . . . . 37 3.4 Scope of Explainable AI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 3.4.1 Stage of Explainability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 3.4.2 Model-dependency of Explainability . . . . . . . . . . . . . . . . . . . . . 39 3.4.3 Scope of Explainability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 3.4.4 Type of Explanations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.4.5 Evaluation of Explainability . . . . . . . . . . . . . . . . . . . . . . . . . . 41 3.5 Review of Explainable AI Methods in Medical Prognosis . . . . . . . . . . . . . . 42 3.5.1 Feature Importance-based Methods . . . . . . . . . . . . . . . . . . . . . . 49 3.5.2 Counterfactual-based Approaches . . . . . . . . . . . . . . . . . . . . . . . 52 3.6 Challenges of Explainable AI Methods in Medical Prognosis . . . . . . . . . . . . 53 3.6.1 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 3.6.2 Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 4 Explainable Breast Cancer Recurrence Prediction based on Clinicopathological Data 55 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 4.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 4.2.1 Computer Aided Prognosis for BC Recurrence Prediction . . . . . . . . . 56 4.2.2 Explainable AI for BC Prognosis . . . . . . . . . . . . . . . . . . . . . . . 56 4.3 Proposed Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 4.3.1 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 4.3.2 Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 4.3.3 Cost-Sensitive CatBoost based Classification . . . . . . . . . . . . . . . . 59 4.3.4 Quantification of Features Impact . . . . . . . . . . . . . . . . . . . . . . 59 4.3.4.1 LIME based Quantification . . . . . . . . . . . . . . . . . . . . . 59 4.3.4.2 SHAP-based Quantification . . . . . . . . . . . . . . . . . . . . . 61 4.4 Experimentation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63ix 4.4.1 Dataset Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 4.4.2 Classification Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 4.4.2.1 Performance with versus without Feature Selection . . . . . . . . 66 4.4.2.2 Performance of Classical versus Boosting Classifiers . . . . . . . 67 4.4.2.3 Performance of Cost-Insensitive versus Cost-sensitive Classifiers 69 4.4.2.4 Optimal Cost Value for Cost-Sensitive Learning . . . . . . . . . 70 4.4.2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 4.4.3 Results of Explainability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 4.4.3.1 LIME Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 4.4.3.2 SHAP Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 4.4.3.3 Comparison of LIME and SHAP Explainability Methods . . . . 75 4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 5 Explainable M-CoGCN for Multimodal Breast Cancer Recurrence Prediction 80 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 5.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 5.2.1 Desired Properties of Counterfactual Explanations . . . . . . . . . . . . . 81 5.2.2 Approaches for Counterfactual Explanations Generation . . . . . . . . . . 82 5.3 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 5.3.1 Graph-based Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 5.3.2 Cross-attention based Multimodality Data Fusion . . . . . . . . . . . . . . 84 5.3.3 Counterfactual Explanations . . . . . . . . . . . . . . . . . . . . . . . . . 85 5.4 Proposed Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 5.4.1 M-CoGCN Model Construction . . . . . . . . . . . . . . . . . . . . . . . . 89 5.4.1.1 Patient Similarity Networks Construction . . . . . . . . . . . . . 89 5.4.1.2 Patient Embedding Generation . . . . . . . . . . . . . . . . . . . 90 5.4.1.3 Cross-Modality based on Co-Attention Mechanism . . . . . . . . 90 5.4.1.4 Graph-aware Fusion Gate . . . . . . . . . . . . . . . . . . . . . . 92 5.4.1.5 Recurrence Prediction . . . . . . . . . . . . . . . . . . . . . . . . 93 5.4.1.6 Loss Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 5.4.2 Counterfactual Explanations for Prognosis Enhancement . . . . . . . . . . 94 5.4.2.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . 94 5.4.2.2 NSGA-II Components and Evolution . . . . . . . . . . . . . . . 95 5.5 Experimentation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 5.5.1 Data Description and Preprocessing . . . . . . . . . . . . . . . . . . . . . 98 5.5.2 Recurrence Prediction Results . . . . . . . . . . . . . . . . . . . . . . . . . 100 5.5.3 Ablation Study on the Number of Edges . . . . . . . . . . . . . . . . . . . 104 5.5.4 Ablation Study on the Number of Co-Attention Heads . . . . . . . . . . . 106 5.5.5 Ablation Study on Fusion Gate . . . . . . . . . . . . . . . . . . . . . . . . 107 5.5.6 Counterfactuals Generation Results . . . . . . . . . . . . . . . . . . . . . 108 5.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110x 6 General Conclusion 112 6.1 Thesis Aims . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 6.2 Thesis Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 6.3 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 List of Publications 116 Bibliography 11 |
| Type de document : | Thése doctorat |
Disponibilité (1)
| Cote | Support | Localisation | Statut |
|---|---|---|---|
| TINF/221 | Théses de doctorat | bibliothèque sciences exactes | Consultable |




