Dana-Farber Cancer Institute investigators and collaborators at Mass General Brigham have created a single algorithm that simultaneously predicts the likelihood of 348 distinct diseases for a given patient. The algorithm, described in Nature, is the first to make multi-disease predictions based only on routinely collected electronic health record data plus knowledge about the patient's genetic risks of disease. The algorithm dynamically updates its risk assessments over time, and its predictive power improves as the patient ages.
"This is a novel algorithm that analyzes full patient clinical data trajectories to make very accurate predictions of disease transitions into the future," says co-senior author Alexander Gusev, PhD, a Dana-Farber scientist. "People are thinking about what their health is going to look like over the next few years, especially with increasing intervention options. This tool offers a path toward improving the prediction of future diseases so doctors and patients can take action to try to prevent them."
The team, led by co-senior authors Gusev, Giovanni Parmigiani, PhD, a Dana-Farber researcher and associate director of the Division of Population Sciences, Pradeep Natarajan of Massachusetts General Hospital (MGH), and first-author Sarah Urbut, MD, PhD, an MGH cardiologist, combined probabilistic modeling with machine learning in a unique design that makes it possible for the tool's predictions to be both powerful and interpretable biologically.
The researchers employed probabilistic modeling to define 20 biological "signatures," or sets of biological trends, that have a high probability of initiating certain diseases. For example, high cholesterol in a patient's history increases the probability of cardiovascular diseases. The signatures also factor in genetic variations known to increase the likelihood of a disease. These signatures are complex and overlapping. The risk of colon cancer, for instance, is associated with multiple different signatures.
"We have painstakingly curated these signatures, which is a big differentiator. In contrast to deep learning approaches, which are typically 'black boxes,' our curated signatures capture the underlying biology in an interpretable way," says Parmigiani. "These signatures are then the drivers of the model's ability to make predictions."
The team used artificial intelligence techniques to train a model to predict the risk of a wide range of diseases based on the patient's health history, viewed through the lens of their curated disease signatures and a learned understanding of the co-occurrence of diseases.
"There is a lot of useful information in a medical record, both over time and across different disease areas," says Parmigiani. "That data would be difficult for a human to process in their head, but tractable for a machine learning model."
The model, called Aladynoulli (a portmanteau blending Aladdin, who had access to wish-granting genies, with the idea of dynamic predictions based on probabilities, as defined by the mathematician Jacob Bernoulli) was trained and validated using three large biobanks including a total of over 683,000 patient records. The model outperformed existing cardiovascular risk models - PCE, QRISK3, and PREVENT - with more accurate 10-year predictions. It also outperformed the GAIL breast cancer risk model for 1-year breast cancer predictions.
The model can also predict which patients will develop colorectal cancer in the coming year with very high accuracy. Flagging a high risk of imminent colorectal cancer could help a primary care physician refer a patient for a colonoscopy, even if that patient is not yet eligible for screening based on current age-based guidelines. Given that young onset colorectal cancer is increasing at an alarming rate, this type of risk assessment could help doctors intervene early, when the disease is still preventable.
"In clinic, two patients with the same diagnosis are not the same patient, said Pradeep Natarajan, MD, director of Preventive Cardiology at Mass General Brigham Heart and Vascular Institute and associate member at the Broad Institute of MIT and Harvard. "This model shows they often have different underlying signature profiles, which can translate into different progression patterns and different responses to the same treatment."
"This model is an innovative and potentially disruptive tool in the clinic because it could encourage more clinicians to think in a cross-disciplinary way," says Parmigiani. "It is important to try to understand a patient with a full 360-degree perspective on the data, as opposed to compartmentalizing the information by specialty. One of the biggest strengths of this model is that it knows that many diseases are driven by the same underlying biology."
Next steps for the team include using the model to learn more about why and how melanoma, a form of skin cancer, spreads to other organs. The team has partnered with Dana-Farber medical oncologist David Liu, MD, MPH, to train the model to identify different patterns of metastases, such as spreading early or late, or spreading widely or narrowly. The work could help identify previously unrecognized subtypes of the disease.
"Aladynoulli not only can help predict different trajectories of metastases, but it also can give us a biological understanding of why patients are progressing along these different trajectories," says Gusev. "From a better understanding of the biology, we have the potential to identify more beneficial therapeutics."
The team is simultaneously working to expand the signatures in the model to increase the accuracy and biological grounding of the risk predictions. They are also looking for opportunities to move towards implementing the model in clinical practice and as a tool to improve the design of clinical trials.
