A Comprehensive, Artificial Intelligence, Digital Twin Platform Based on Multimodal Real-World Data Integration for Personalized Medicine in Hematology

Background. Personalized medicine in hematology requires extensive real-world and comprehensive data, including clinical and genomic information. However, integrating, processing and managing such complex data layers in large populations presents significant challenges. Development of patient-tailored models by Artificial Intelligence (AI), known as Digital Twins (DT) offers a novel approach to precision medicine. DT are virtual representations of patients created from multimodal information that can be used to improve diagnosis, prognosis and treatment outcome, improving clinical decision-making. This project aims to advance research by using AI to develop a DT platform for personalized medicine in hematology, with myelodysplastic syndromes (MDS) as case study. MDS are hematological diseases with high clinical and genomic heterogeneity, presenting a challenging scenario for new technologies implementation

Methods. We exploited the Federated Learning platform and technology for Synthetic Data generation developed by the GenoMed4All and Synthema consortia to collect multimodal data from a broad population of MDS patients, in a privacy-complaint manner avoiding data sharing. The multimodal longitudinal data included both structured and unstructured information, encompassing clinical records, genomics, image data, treatments, longitudinal outcomes, and patients reported outcomes. We analyzed 22,080 MDS patients by gathering data using different technological strategies from retrospective cohorts of the international consortium for MDS, prospective Italian MDS registry (FISIM), GIMEMA clinical network, Humanitas Hospital DataLake and Electronic Health Records form the central repository of Lombardia region, Italy
Clinical, genomic, and imaging data were organized and integrated into a comprehensive DataLake architecture in common data model format. A Retrieval-Augmented Generation (RAG) system with a pre-trained Large Language Model (LLM) was implemented to search, extract and summarize patient information across multiple unstructured text documents and medical records. An innovative AI-based tool was used to capture information on Patient-Reported Outcomes. SHapley Additive exPlanations (SHAP) approach was utilized to explain features importance and impact on model prediction. Several AI models were integrated into an interactive and comprehensive DT platform to address relevant clinical questions in MDS patients

Results. The DT platform (GEMINI) enables manual entry of individual patient information, including age, sex, blood parameters, gene mutations, cytogenetic abnormalities, morphological and histological features along with clinical questions to address (diagnosis, prognostic assessment, and treatment strategy). First, the DT model infers the conditional dependencies among mutations and classifies individual patients into specific clusters, capturing broad dependencies among genomic alterations. Then, using an AI framework for multimodal analysis in cancer (based on PMID35944502), the DT platform provides personalized estimates of 1) probability of survival, 2) leukemic evolution risk, and 3) response to specific treatments. Using individual clinical and genomic features within a multi-state Markov model, the DT serves as a decision support system to define the optimal timing for therapeutic interventions, focusing on transplantation. Finally, for each disease state and clinical scenario, the DT can simulate changes in quality of life and patient-reported outcomes. This platform, intended for research purposes only, is publicly available (https://gemini-xkb3corsxq-ew.a.run.app/) to help clinicians become familiar with the DT concept. An interactive interface allows users to explore a high-fidelity simulation of the disease’s natural history, the impact and timing of therapeutic interventions based on individual patient characteristics

Conclusion. GEMINI offers a robust system that supports clinical decision-making, providing a comprehensive overview of the natural history of the disease and the patient's perspective. Our solution benefits from the integration of multiple sources of information in a privacy-compliant manner, avoiding data sharing. DT are expected to improve and accelerate personalized medicine approaches across different hematological diseases with unmet clinical needs