Doctors at Massachusetts General Hospital and Dana-Farber say they have built a new artificial-intelligence system that combs through routine medical records, plus genetic test results when they are available, to estimate a patient’s chances of developing any of 348 different conditions. The hope is to plug the tool into electronic health records so it can quietly flag patients who might need earlier screening or closer follow-up.
How the model works
The algorithm, called Aladynoulli, uses a Bayesian, signature-based approach that tracks how disease risks shift over time and across specialties. Each new visit or lab result updates the picture, so the prediction is not frozen at a single point in a patient’s life.
As detailed in the medRxiv preprint and the project documentation, the team trained Aladynoulli on about 683,000 patient records pulled from UK Biobank, Mass General Brigham and the All of Us cohort, adding genetic data when it was on hand. The authors also released code and interactive notebooks so reviewers and clinicians can see how the system arrives at its predictions instead of treating it like a black box.
Researchers say it "sees the patient holistically"
“It sees the patient holistically, both across departments and over time,” Dr. Sarah Urbut, an MGH cardiologist and the study’s lead author, told The Boston Globe. Senior author Giovanni Parmigiani told The Globe that the model can recognize combinations of diagnoses and patterns that would be very hard for a human clinician to juggle mentally.
The paper’s title, Aladynoulli, tips its hat to both Aladdin and the Bernoulli family of probability, according to The Globe. The outlet also reports that the group discussed publishing the work in a Nature journal as part of the rollout.
Outperforms common risk scores
In head-to-head testing, the team reports that Aladynoulli outperformed commonly used online calculators for several tasks, including 10-year cardiovascular risk scores and short-horizon breast cancer risk estimates, according to the preprint. Instead of relying on a single clinical snapshot, the system leans on longitudinal records.
The authors say that by modeling hundreds of conditions together, the system can “borrow” signal from related diagnoses, which in turn improves prediction for less common outcomes.
Signatures point to specific risks
The researchers compressed patient histories into about 20 biological “signatures” and then identified clusters of conditions linked to particular cancers, inflammatory problems and metabolic issues. The Boston Globe notes that one cluster included multiple colorectal-related conditions, such as ulcerative colitis and other gastrointestinal complications.
The model was especially strong at flagging people likely to develop colorectal cancer within a year, the Globe reports, a capability that could nudge clinicians to order screenings sooner for high-risk patients.
Screening, equity and regulation
That kind of early signal matters because colorectal cancer cases and deaths have been rising among younger adults. The American Cancer Society has highlighted the shift in age patterns and emphasized the importance of timely screening.
Any real-world deployment will also have to satisfy federal rules. The FDA treats many clinical decision-support algorithms as software-as-a-medical-device and has issued guidance on testing, transparency and lifecycle monitoring for AI tools.
The investigators say Aladynoulli is ready for careful clinical trials, but that widespread hospital use will require prospective validation, checks for bias and a clear regulatory strategy. For now, the team’s code and interactive notebooks stay public for clinicians and health systems that want to study the model before any potential rollout.
