ADAM Substudy Luik 2: Observational Retrospective Single Centre Cohort Study on Constructing & Validating AI Prediction Models for Outcomes of Lung Cancer Patients

The aim of this study is to prepare and unlock siloed real-world-data (RWD) for analysis and artificial model construction with the goal to derive real-world-evidence (RWE). Nowadays physicians and nurses are gathering data from patients and register the data in the electronic health record system. Datapoints are often not available in a structured format, so data gathering and unlocking is done manually upon request. Careful analysis of the collected data using artificial intelligence tools might also help to predict which patients are at the highest risk of unscheduled health care use, emergency department visits and hospital admissions. Therefore, a model able to predict the relevant outcomes would be of significant help to the physicians' daily practice.

Primary Objective Developing and validating an AI model that supports physicians in their decision process for treating lung cancer patients. This AI model needs to predict the probability (of the evolution) of the outcomes, based on clinical data and a simulated lung cancer treatment plan and later on a deep patient. Model input data sources are hospital data like demographics, baseline health status, prior treatments, tumour characteristics, comorbidities , imaging data & physiological data, detailed treatment data.

Secondary Objectives

• Automatic unlocking, collection & transformation of lung cancer datapoints to OMOP common data model so that data is readily available for further research & analysis
• Training and validating supervised machine learning models with a limited feature set as input to predict lung cancer patient outcomes
• Constructing a digital patient by training an AI model fed with all data available in OMOP common data model
• Validating a digital patient & optimal feature selection to enhance AI model performance via unsupervised learning techniques

The potential of applying transformers to represent patients is truly personalized and even predictive medicine. The reason is that transformers are an instrument which make it possible to deal with millions of interacting and non-linearly behaving parameters. Hence data sources can be extended to include genetic information and so on. Optimal feature selection from a digital patient can enhance AI model performance via unsupervised learning techniques, and so further finetune prediction models for daily practice.