AI Model Revolutionizes Heart Disease Risk Prediction, Adapting to Real-World Clinical Challenges
Researchers at ETH Zurich and the Max Planck Institute for Intelligent Systems have created a novel artificial intelligence model that might revolutionize patient risk of cardiovascular disease assessment by doctors. Called AdaCVD, the new model shows an unheard-of capacity to adapt to the messy reality of clinical practice and flexibly include various patient data.
Claiming almost 18 million deaths annually, cardiovascular disease (CVD) continues to be the main cause of death globally. Starting early preventive actions depends on precisely determining each person's CVD risk. But current risk assessment instruments have major shortcomings that limit their usefulness in clinical environments.
Lead researcher Dr. Frederike Lübeck says "current models struggle with incomplete or inconsistently formatted data and rely on rigid sets of input variables." "But patient information is often messy, unstructured, and varies greatly between different healthcare environments in actual clinical practice."
Using large language models (LLMs), the same kind of artificial intelligence technology behind chatbots like ChatGPT, the research team sought to overcome these obstacles. Using almost 500,000 UK Biobank study participants, they carefully tuned an LLM on comprehensive health data.
AdaCVD, the resulting model, matched specialized machine learning techniques in forecasting 10-year CVD risk and exceeded conventional risk assessments. Its actual strength, though, is in its adaptability across three important dimensions that mirror actual clinical complexity:
Incorporating flexible patient information.
Handling unstructured text like clinical notes in addition to organized data.
Quickly adjusting to new patient demographics with little more information.
"What distinguishes AdaCVD is its capacity to reason over many patient representations, from neatly structured data tables to free-form clinical notes," co-author Dr. Jonas Wildberger says. "It can extract relevant information from whatever patient data is available, without being limited by rigid input requirements."
AdaCVD attained state-of- the-art performance with an area under the receiver operating characteristic curve (AUROC) of 0.738 in benchmark comparisons using only basic risk factors including age, cholesterol levels, and blood pressure. This matched the top-performing machine learning models made especially for structured inputs and exceeded accepted medical risk scores.
However, the model really excels when including more general patient data than only the usual risk factors. AdaCVD's predictive performance increased by almost 5%, reaching an AUROC of 0.774, when given access to complete data on lifestyle, family history, lab tests, and more.
"We discovered that risk assessment is much improved by combining many facets of a patient's health profile," says Dr. Lübeck. "Factors including detailed lifestyle information, genetic risk scores, and past medical history all contribute valuable predictive signals."
To evaluate possible biases, the researchers ran thorough analyses over several demographic, socioeconomic, and clinical subgroups. They discovered that including thorough medical knowledge into risk prediction helped every subgroup.
Especially some underrepresented groups observed much more performance improvement. Elderly people (7.55% increase), those without higher education (7.50% increase), current smokers (7.46% increase), and those with diabetes (7.21% increase) showed the biggest improvements.
"This suggests that including thorough health information may be especially beneficial for identifying high-risk individuals in traditionally underprivileged populations," notes Dr. Wildberger. "It could help to lessen differences in preventative care."
AdaCVD's ability to manage incomplete or variable patient information is a major breakthrough since this is a typical problem in clinical environments. The researchers contrasted several methods for handling missing data.
They discovered that additional fine-tuning the model on patient descriptions with randomly varying degrees of detail enabled it to generate accurate predictions from any accessible data. Dubbed AdaCVD-Flex, this adaptable form kept strong performance even with partial inputs.
"AdaCVD-Flex can work with whatever patient data is available rather than requiring a fixed set of variables," notes Dr. Lübeck. "This makes it much more suitable for deployment across various healthcare environments with different data collecting methods."
The model also shown a remarkable capacity to reason over unstructured text inputs such as clinical notes, a common data format in actual medical environments. AdaCVD could be effectively adjusted to process free-text patient summaries with minimum additional fine-tuning following first training on structured data.
Comparatively to models trained from scratch, this text processing capacity needed 100 times less training data. It creates great opportunities for using rich clinical data without much data preparation.
The researchers tested the model on data from the Framingham Heart Study, a classic cardiovascular research initiative carried out in the United States beginning in the 1940s, to assess its capacity to generalize to new populations. When compared to the UK Biobank data used for first training, this marked a notable change in both time period and geographical area.
AdaCVD obtained good zero-shot performance on the Framingham dataset even without any further fine-tuning. Minimal fine-tuning on a small subset of Framingham data allowed it to rapidly adapt to match the performance of the original Framingham risk score, derived from the whole dataset.
"This shows AdaCVD's fast adaptability to new patient populations using minimum additional data," says Dr. Wildberger. "It could help knowledge flow between several healthcare settings to be more effective."
The researchers underline that their work is meant to enhance human clinical judgement with AI-powered risk stratification, not to replace it.
"The objective is to give clinicians a flexible tool that can integrate whatever patient information is available to identify high-risk individuals who may benefit from early interventions," says Dr. Lübeck. "This could help direct more customized preventive plans."
The team notes some limits of their research even if the results show promise. The model was mostly trained and tested on UK data; more validation across more varied worldwide populations is therefore necessary. Using synthetic generated clinical notes, they also emphasized the need of high-quality datasets including real clinical text matched with long-term health outcomes.
Still, this work marks a significant first towards more adaptable and strong artificial intelligence systems for clinical decision support. AdaCVD presents a viable route for using machine learning in messy real-world healthcare environments by tackling important challenges around data heterogeneity and distribution changes.
"We're thrilled about the possibility for this technology to enhance cardiovascular risk assessment and ultimately improve patient outcomes," notes Dr. Lübeck. "But it's crucial that such tools be rigorously validated and carefully integrated into clinical workflows in partnership with healthcare providers."
Research in this field may bring in a new era of AI-augmented preventive care whereby intelligent systems can quickly adapt to local settings using insights from many worldwide data sources.
For millions of people who run the danger of heart disease, these developments could literally save lives.
The researchers have made their code publicly available to facilitate further research in this important area.