Cardiovascular

Survival Prediction of TAVI Patients Using End-to-End Deep Image Regression on Whole-Body CT.

Müller D, Rizoudis A, et al. • Studies in health technology and informatics • 2026

TL;DR

An end-to-end deep learning pipeline using 3D DenseNet121 on whole-body CT scans achieved an AUC of 0.787 for predicting short-term survival in TAVI patients, comparable to state-of-the-art multimodal ML models requiring extensive clinical data.

Key Findings

Results

A 3D DenseNet121 deep learning model trained on whole-body CT scans achieved an AUC of 0.787 for identifying short-term survival (less than 12 months) in TAVI patients.

The model was trained using 5-fold cross-validation on 317 TAVI patients
Binary evaluation threshold was set at 12 months survival
AUC of 0.787 was achieved in the binary classification evaluation
The model operated as a regression model predicting survival in days before binary thresholding was applied

Results

The deep learning model achieved a specificity of 72.1% and a sensitivity of 85.3% for identifying short-term survival patients.

Specificity of 72.1% was achieved for identifying patients with survival less than 12 months
Sensitivity of 85.3% was achieved for the same binary classification task
These metrics were derived from 5-fold cross-validation on 317 TAVI patients
The high sensitivity suggests the model is particularly effective at identifying high-risk patients who will not survive beyond 12 months

Results

The end-to-end deep learning pipeline performed comparably to state-of-the-art multimodal machine learning models that require extensive clinical data.

The pipeline operates directly on routinely acquired pre-procedural whole-body CT scans without manual feature engineering
Traditional risk scores such as EuroSCORE II and STS-PROM exhibit poor calibration for TAVI cohorts
The model uses the AUCMEDI framework with a 3D DenseNet121 architecture
Performance was achieved without requiring the extensive clinical data inputs needed by competing multimodal ML models

Background

Traditional risk scores EuroSCORE II and STS-PROM show poor calibration for TAVI patient cohorts, motivating the development of alternative prediction methods.

EuroSCORE II and STS-PROM are identified as standard traditional risk scores for this patient population
Both scores exhibit poor calibration specifically for TAVI cohorts
This limitation was a primary motivation for developing the deep learning approach
Predicting long-term survival outcomes in TAVI patients is described as 'challenging'

Methods

The study cohort consisted of 317 TAVI patients whose pre-procedural whole-body CT scans were used for model training and evaluation.

317 TAVI patients were included in the study
CT scans were pre-procedural, meaning they were acquired before the TAVI procedure
5-fold cross-validation was used for model training and evaluation
CT scans were described as 'routinely acquired,' indicating no special imaging protocols were required

Methods

The deep learning pipeline was designed as an end-to-end regression model predicting survival in days directly from CT images.

The model was trained as a regression model outputting survival prediction in days
No manual feature engineering was required
The pipeline used a 3D DenseNet121 architecture within the AUCMEDI framework
The end-to-end approach processes whole-body CT scans directly without intermediate segmentation or feature extraction steps
The regression output was converted to binary classification using a 12-month threshold for evaluation

What This Means

This research suggests that artificial intelligence can analyze routine CT scans taken before a heart valve procedure called TAVI (transcatheter aortic valve implantation) to predict which patients are at high risk of dying within one year. The researchers trained a deep learning model on CT scans from 317 patients, and the model was able to identify short-term survivors with an accuracy measure (AUC) of 0.787, correctly flagging about 85% of high-risk patients. Importantly, this was achieved without doctors or engineers having to manually extract measurements or clinical features from the scans — the AI learned directly from the raw images. This matters because TAVI is increasingly used to treat elderly or frail patients with severe heart valve disease, but current tools used to predict risk (called EuroSCORE II and STS-PROM) do not work well for this specific group of patients. Having a better way to identify which patients are unlikely to benefit from the procedure could help doctors and patients make more informed decisions about treatment. The fact that this AI model performed as well as more complex systems that require extensive clinical data — using only the CT scan that is already routinely performed before the procedure — makes it particularly practical for real-world clinical use. This research suggests that incorporating such a tool into pre-operative assessment could help identify high-risk TAVI patients without adding extra tests or burdens to clinical workflows. However, the study was conducted on a relatively small cohort of 317 patients, and further validation on larger and more diverse patient populations would be needed before such a tool could be used in clinical practice.

Have a question about this study?

Citation

M&#xfc;ller D, Rizoudis A, Lind C, Kramer F, Voran J, Ulrich H. (2026). Survival Prediction of TAVI Patients Using End-to-End Deep Image Regression on Whole-Body CT.. Studies in health technology and informatics. https://doi.org/10.3233/SHTI261005

Key Findings

What This Means

Have a question about this study?

Related Research

Citation