AI Spots Frailty and Mortality Risk from a Single Chest X-Ray

Patients flagged as frail had nearly an eightfold higher risk of mortality, according to research presented Monday.

Although the consequences of frailty can be fatal, assessments proven to help identify the syndrome and predict adverse outcomes remain underutilized. A new deep learning AI model puts these measures to work, revealing signs of frailty in routine chest X-rays and demonstrating early promise as an opportunistic screening tool.

Researchers in South Korea who developed the AI tool aim to help clinicians intervene earlier and improve quality of life and clinical outcomes for older adults. “Routine X-rays contain valuable, hidden data about a patient's overall frailty risk,” said study co-author Namkug Kim, PhD, professor at the University of Ulsan College of Medicine in Korea.

Dr. Kim presented on behalf of Seungjoo Park, BS, a medical student at CHA University and study lead author, who was unable to attend RSNA 2025 in person. “Our validation of this AI model is a step toward an automated screening system that, in the future, could act as an early warning, automatically flagging high-risk patients whose visits to the doctor or ER may be for unrelated reasons but they would benefit from a full comprehensive geriatric assessment,” Dr. Kim said.

Finding Hidden Value in Medical Data

The multitask AI tool leverages the power of the gold standard Frailty Index (FI), a complex and objective metric calculated from the accumulation of dozens of health deficits. According to Dr. Kim, a pre-training approach using less elaborate proxy tasks, including age, sex and the nine-point Clinical Frailty Scale (CFS), was a critical step in building a model powerful enough to successfully predict the FI.

The researchers leveraged three distinct datasets from the Asan Medical Center, a tertiary hospital in Korea to build and validate the AI tool. First, they used 11,987 chest X-rays from patients aged 65 and older, each accompanied by CFS records to train the model to recognize signs of frailty. Next, they refined the model using 1,359 chest X-rays from patients who had received comprehensive geriatric assessments, including FI.

To test the model’s reliability, the team used the records of 5,932 individuals aged 65 and older from a health screening cohort whose screening records included bioelectrical impedance analysis, chest X-ray and survival information.

The researchers tested a model trained to spot frailty using the FI alone, but its accuracy was limited as they expected given the complexity of the index, the relative scarcity of F1 data, and the benefits of multi-task learning.

When they pretrained the model with additional information like age, sex and scores from the CFS, then fine-tuned it using F1 data, the model became much better at identifying frail patients. It could also correctly spot non-frail patients.

“AI-predicted frailty status is strongly associated with long-term patient survival,” Dr. Kim said. “Patients flagged as frail had a 7.7 times greater risk of mortality.”

The model showed a significant correlation with biological age, predicting a person’s age within about four years of the true biological age. The AI tool was also highly accurate at predicting each person’s sex.

“This research is a prime example of how AI can help unlock the hidden value in medical data we are already collecting, shifting patient care from reactive to preventative,” Dr. Kim said. “We are hopeful this technology can make a real impact in clinical practice for our aging society and we invite and welcome contact from institutions interested in collaboration or externally validating our model.”

Access the scientific poster, “Opportunistic Screening of Frailty from Chest Radiographs and Predicting Survival in Elderly,” (M2-SPCH-7) on demand at RSNA.org/MeetingCentral.