CHICAGO — A 3D convolutional neural network (CNN) model that uses chest CT data predicts patients’ risk of major adverse cardiovascular events (MACE), according to research presented December 4 at the RSNA meeting.
And the algorithm appears effective even when applied to diverse patient groups, according to presenter Amara Tariq, PhD, of the Mayo Clinic College of Medicine and Science in Rochester, MN.
“[The] model outperformed comparative models and generalized well to external data with significant shift in patient populations,” Tariq told session attendees.
Traditionally, MACE risk assessment is done using clinical and demographic factors or through coronary artery calcium (CAC) scoring on CT. But assessing risk of major adverse cardiovascular events (MACE) in these ways can be limited, Tariq said — especially for minority patients.
“Standardized risk scores such as PREVENT or PCE [Pooled Cohort Equations] don’t include imaging and are focused on known risk factors such as age or gender,” she said. “[These techniques have shown] suboptimal performance for minority subgroups. As well, CAC scores may ignore broader health implications, such as body mass index.”
Tariq and colleagues sought to address the problem by using AI for MACE risk assessment with non-gated, non-contrast chest CT exams performed for reasons other than cardiac evaluation. This way, gauging MACE risk isn’t limited to known risk factors only, and makes evaluating heart disease risk equitable “across diverse populations,” she said.
The team’s study included 4,431 noncontrast chest CT exams performed between 2015 and 2022 at three Mayo Clinic sites; it used these exams as an internal test set to develop the AI algorithm. The group also included a separate cohort of 201 chest CT exams taken from Emory University Hospital in Atlanta for validation of the algorithm. This last group included more Black patients (29% compared with 10%) and more individuals with comorbidities such as diabetes, hypertension, and chronic kidney disease.
“The racial distribution was very different among our two cohorts, which gave us a good opportunity to test our generalization and algorithmic debiasing techniques,” she said.
The group trained the model with 3D sub-volume heart region data from the CT exams, which it then used to predict MACE risk over five years.
The investigators reported that overall, the 3D CNN model improved sensitivity and specificity compared with PREVENT assessment and AI-based CAC scoring in both the Mayo Clinic “holdout” set and the Emory external set.
|
Performance comparison for predicting MACE events over 5 years |
||||
|
Test set |
PREVENT |
AI-based CAC scoring |
3D CNN model
|
3D CNN with debiasing |
|
Mayo Clinic |
||||
| Sensitivity |
60.4% |
51.1% |
64.2% |
66.1% |
| Specificity |
53.5% |
58.2% |
74.2% |
68.1% |
| AUROC |
0.56 |
0.57 |
0.73 |
0.72 |
|
Emory |
||||
| Sensitivity |
41.9% |
66.7% |
58% |
70% |
| Specificity |
71% |
40.9% |
73.3% |
62.4% |
| AUROC |
0.54 |
0.52 |
0.65 |
0.69 |
Tariq noted that within the Mayo Clinic test set, the AUROC difference between the two 3D CNN models was not significant, but within the Emory test set — which included a more diverse set of patients — the AUROC was significantly better, suggesting that the 3D model with debiasing was generalized more successfully in this cohort.
Going forward, Tariq and colleagues plan to develop multimodal models that would include CT exam data, radiology reports and clinical notes, and electronic health information.
“Leveraging AI to analyze rich, underutilized data within routinely performed chest CTs holds significant promise for enhancing early MACE risk estimation across diverse patient populations beyond the capabilities of current clinical tools,” she concluded.
To see full coverage of RSNA 2025, visit our RADCast.