A multi-modal AI model combining clinical features and HRCT image embeddings significantly improved the prediction of future acute exacerbations in COPD patients compared to clinical features alone.
Observational
Does a multi-modal AI model combining HRCT image embeddings and clinical features improve the prediction of acute exacerbations in COPD patients compared to clinical features alone?
Combining HRCT image embeddings with clinical data in a multi-modal AI model improves the prediction of acute exacerbations in COPD patients, potentially aiding in identifying high-risk patients for clinical trials.
Abstract Rationale Most clinical trials assessing treatments for acute exacerbation (AEx) in patients with COPD enrich the study cohort for patients likely to experience future AEx based on history of AEx, severity of airflow limitation and patient symptoms. Recent publications suggest HRCT scans contain imaging features that can predict future AEx events 1 and thus identify patients at high risk for AEx for clinical trials. The aim of this work is to retrospectively use the COPDGene data set to explore machine learning models using either HRCT data alone, or in combination with clinical features, and compare the performance of these models to a benchmark prognostic model based on a subset of clinical features alone. Methods For the benchmark prognostic model of AEx, we selected a subset of clinical features (history of AEx, age, gender, smoking status) using the Gini feature importance (XGBoost model) and trained a logistic regression model to predict occurrence of moderate to severe exacerbations at follow up visits for the COPDGene patients. We used Merlin, a vision language Foundation Model (FM) 2 trained on over 15000 3D full body CT scans, electronic health record (EHR) data and extracted image embeddings from HRCT scans of COPDGene patients at baseline and built a multi-modal model combining in a dimension-balanced manner a set of selected clinical features with image embeddings for prediction of future exacerbation. We use a five fold cross validation approach to compare results of logistic regression models trained on image embedding, clinical features only as well as when the two modalities are combined. Results Results are summarized in table 1. Results suggest that each data modality contributes valuable information that are relevant for predicting future AEx. While the clinical features alone perform better than the image based embeddings extracted from the Merlin foundation model without further fine tuning, the combined multi-modal AI model performance significantly improves prediction. Conclusions Each separate modality (clinical/tabular data and HRCT) of patient data contains valuable information that can predict future AEx in COPD patients. The multi-modal AI model demonstrates that image embeddings extracted from HRCT scans contribute information to improve the prognosis of AEx in COPD patients. In the future, we plan to further fine tune and validate the results using additional independent datasets such as SPIROMICS. This abstract is funded by: Genentech inc
Mirsharif et al. (Fri,) conducted a observational in COPD. Multi-modal AI model (clinical features + HRCT image embeddings) vs. Clinical features alone (benchmark prognostic model) was evaluated on Occurrence of moderate to severe exacerbations at follow up visits. A multi-modal AI model combining clinical features and HRCT image embeddings significantly improved the prediction of future acute exacerbations in COPD patients compared to clinical features alone.