Validation of multimodal deep learning models for the prediction of risks of lung cancer and adverse cardiovascular events using chest radiographs and clinical data

Principal Investigator

Name
EUIJIN HWANG

Degrees
MD, PhD

Institution
Seoul National University Hospital

Position Title
Associate Professor

Email
ken92002@snu.ac.kr

About this CDAS Project

Study
PLCO (Learn more about this study)

Project ID
PLCOI-1958

Initial CDAS Request Approval
Nov 24, 2025

Title
Validation of multimodal deep learning models for the prediction of risks of lung cancer and adverse cardiovascular events using chest radiographs and clinical data

Summary
Chest radiograpy is the most commonly used imaing study for screening and diagnosis of various diseases of lung and heart. Multiple studies reported that deep learning-based artificial intelligence models can predict the risk of various diseases such as lung cancer and cardiovascular diseases from chest radiographic images. Although chest radiograph may contain information regading the risk of cardiopulmonary diseases, predictions solely based on chest radiograph may have limitations because various non-image infomation (e.g., age, biological sex, socioeconomical information, family history of disease, individual medical history) cannot be integrated in the prediction. Transformer-based deep learning models have shown great potential in integrating diverse data types, and this multimodal-based prediction may enhance the precision of predicting risk of diseases.
We have developed deep learning algorithms to predict the risks of lung cancer and adverse cardiovascular events (myocardial infarction and cerebrovascular accident) from chest radiographs of healthy individuals and their non-image risk factors, using a dataset from our own institution. We aim to evaluate the performance of developed prediction algorithms using the PLCO trial dataset.

Aims

1. We aim to evaluate the performance of a transformer-based multimodal deep learning model for the lung cancer risk usng the PLCO dataset. The prediction model was developed using dataset of our own institution. The inputs of the prediction model includes chest radiographic image, age, biological sex, body mass index, personal history of cancer or chronic obstructive pulmonary disease, smoking history, and familial history of lung cancer. The model was designed to provide the risk of lung cancer incidence within six years.
2. We aim to evaluate the performance of a transformer-based multimodal deep learning model for the lung cancer risk usng the PLCO dataset. The prediction model was developed using dataset of our own institution. The inputs of the prediction model includes chest radiographic image, age, biological sex, body mass index, personal history of hypertension or diabetes, and smoking history. The model was designed to provide the risk of major adverse cardiovascular events (non-fatal myocardial infarction, non-fatal cerebrovascular accident, and any cardiovascular mortality) within ten years. Since the PLCO dataset does not contain information regarding the incidence of cardiovascular events, we aim to evaluate the performance of the model for the prediction of any mortality due to cardiovascular diseases.

Collaborators

Changi Kim, Seoul National University Hospital
Chang Min Park, Seoul National University Hospital