Dissertation · Nazarbayev University Repository · May 27, 2026

Explainable AI for Clinical Decision Support

Akylbek Oralov, Alisher Kenzhegaliyev, Assylbek Khalyk, Danila Lissovoy, Rakhat Mustafin

Abstract

Robotic fingertips that can grasp delicate objects without crushing them, yet still slide over a surface during tactile exploration, remain a long-standing problem in dexterous manipulation. Syrymova et al. (2025) addressed this with a single-piece silicone fingertip that combines a low-friction hemispherical dome and a high-friction ridged halo, and reported that the device sustains five 50 g batches inside a plastic cup at a closure threshold of c1 = 2 N before gross slip. The present work reproduces that experiment numerically. A simplified solid model of the fingertip is imported from STEP into the MATLAB R2025b PDE Toolbox, support and load faces are mapped explicitly, and a force-controlled compression at 2 N is solved on four progressively refined tetrahedral meshes up to ∼1.9 × 105 elements. A regime-switching Coulomb friction law, with µlow = 0.55 below 1 N and µhigh = 0.72 above, is then driven through a discrete-time simulation of Experiment 1: cup grasp followed by sequential addition of 50 g batches every 10 s. Static FEA produces a peak displacement of 3.68 mm and a peak von Mises stress of 90.4 kPa at the finest mesh, and the dynamic simulation predicts gross slip at the fifth batch, agreeing with the published measurement. The static and dynamic results together establish a baseline that future work can extend with a hyperelastic constitutive law and a fully coupled contact formulation. This report describes the design, implementation, and evaluation of Explainable AI for Clinical Decision Support, a multi-modal clinical decision support platform. The system addresses the practical gap between predictive accuracy and clinical usability: a model that outputs a probability without explanation is difficult for a clinician to validate, trust, or appropriately override. The platform integrates three complementary prediction pipelines operating on different clinical data modalities. A chest X-ray screening module based on a pre-trained CheXNet convolutional neural network achieves a mean AUROC of 0.788 across 14 pathologies evaluated on 7,818 images. A 10-year coronary heart disease risk predictor using a support vector machine trained on Framingham Heart Study data achieves ROC-AUC = 0.704. A 30-day hospital readmission risk predictor using LightGBM trained on the UCI Diabetes 130-US Hospitals dataset achieves ROC-AUC = 0.724. All three pipelines expose formal explainability outputs alongside predictions: Grad-CAM and Grad-CAM++ spatial heatmaps for the imaging module, and SHAP combined with LIME feature attribution for both tabular models. A cross-method agreement count between SHAP and LIME is surfaced as an inference-time reliability indicator. The platform is deployed as a browser-accessible web application backed by FastAPI, PostgreSQL, Redis, and a React frontend, orchestrated via Docker Compose.

Explainable Artificial Intelligence (XAI)Artificial Intelligence in Healthcare and EducationArtificial Intelligence in Healthcareexplainable AIclinical decision supportdexterous manipulation

Explainable AI for Clinical Decision Support

Abstract

Related publications