A new artificial intelligence (AI) model can detect acute heart failure on chest computed tomography (CT) scans of patients hospitalised with shortness of breath. The model performs similarly to both radiologists and cardiologists and may, in the long term, help doctors make better use of the scans that many patients already undergo.
More than one in 20 patients hospitalised in Denmark with an acute condition has acute heart failure – a condition for which rapid treatment can be crucial. The problem is that the symptoms often resemble those of pneumonia, chronic obstructive pulmonary disease and blood clots in the lungs, even though the treatment is completely different. This makes diagnosis difficult, precisely when time is of the essence.
Now, a new study shows that AI can recognise the condition in chest CT scans with an area under the receiver operating characteristic (AUROC) curve of 0.95. AUROC measures how well a test can distinguish between sick and healthy patients, and 1.0 corresponds to a perfect diagnosis.
This is similar to what trained radiologists can achieve.
“We saw an obvious opportunity to use the CT scans that many patients undergo anyway more actively as a diagnostic tool. By letting AI analyse the images, we can potentially detect heart failure earlier and at the same time help give priority to those who need specialist cardiac diagnostics,” explains Kristina Cecilia Miger, a doctor and researcher at the Department of Cardiology at Bispebjerg and Frederiksberg Hospital, Denmark.
The research has been published in European Radiology Experimental.
The signs of heart failure are already visible on the scan
New figures show that heart failure causes 5–7% of all acute hospitalisations and is one of the most common causes among people older than 65 years.
The challenge is that lung diseases or blood clots in the lungs can also cause shortness of breath. Therefore, doctors often have to make decisions before the picture is entirely clear, even though rapid treatment can be crucial to patient outcome.
“The most accurate method is a heart ultrasound, but the examination requires specially trained staff and is not always immediately available. X-rays and lung ultrasounds can help but cannot on their own confirm or rule out the diagnosis with certainty. This means that some patients are either diagnosed too late or initially receive a different diagnosis,” explains Kristina Cecilia Miger.
The AI model has been trained to recognise patterns on CT scans that are typical in acute heart failure – including fluid in and around the lungs, which occurs when the heart is not pumping blood efficiently enough through the body.
When the model was tested against both the standard radiology report, blinded expert radiologists and the cardiologists’ assessments, the AUROC ranged between 0.91 and 0.96. This suggests that the results do not depend on any one specific method of diagnosis.
“We had, of course, hoped for good results, but we were pleasantly surprised by how robust the model performs in comparison with experienced cardiologists and radiologists. The most important thing for us is not just the high AUROC but that it performs consistently with the patients we see in the clinic,” says Kristina Cecilia Miger.
Two thresholds to help confirm or rule out the diagnosis
The study uses three thresholds for the AI assessment, two of which are of particular clinical significance. One threshold is the limit that determines when the model assesses that the risk of heart failure is high enough for the doctor to act.
At the lowest threshold, the model identifies 97% of patients with acute heart failure, thereby missing only a very small number of cases. It can therefore be used in particular to rule out the diagnosis.
At the highest threshold, the model achieves a specificity of 96%, meaning that relatively few healthy patients are incorrectly classified as ill. It can therefore help to confirm the diagnosis and set priorities for further investigation and treatment.
“The idea is that the model can be used in the same way as other diagnostic tests in emergency medicine. A low threshold can help to quickly rule out heart failure for patients with a low probability, whereas a high threshold can identify those for whom the diagnosis is highly likely. It is not about replacing clinical judgement but about supporting decision-making, particularly at night, when specialist radiological or cardiological expertise is not always immediately available,” explains Kristina Cecilia Miger.
Can these hidden clues be used in practice?
The study also showed that patients classified by the AI as having acute heart failure had more than double the mortality rate over a two-year follow-up period. The result suggests that the model largely identifies the patients who are generally the most seriously ill.
However, this association disappeared after adjusting for factors such as age and sex, suggesting that the scan results primarily overlap with known risk markers.
The study was conducted at a single hospital and included only patients aged 50 years or older who underwent low-dose CT scans without contrast. The researchers therefore do not yet know whether the results would be equally good at other hospitals or in other patient groups. The tool must therefore be validated more broadly before it can be put into daily clinical use.
“The next step is larger multicentre validations and studies of how the tool affects clinical decisions and patient outcomes in practice. Furthermore, implementation requires regulatory approval and integration into hospital information technology systems. If the results can be replicated in larger studies, this type of AI decision support is likely to become part of clinical practice over the coming years,” says Kristina Cecilia Miger.
