Grant: Project Grant

When a patient asks, "What is likely to happen to me in the next year?", doctors often base their answer on a snapshot of current information, such as age, sex, whether the person has diabetes, and the most recent lab results like estimated glomerular filtration rate (eGFR) for kidney function. But in most cases, this is not the full story. Electronic health records (EHRs) contain repeated measures over time, such as years of eGFR readings, blood sugar levels (A1C), blood pressure, and medication history. These repeated measures reflect how a person's health has changed over time and may contain important clues about what might happen next. Unfortunately, most current methods for analyzing health data either ignore these repeated measures or simplify them too much, using only the last value or assuming straight-line trends. These simplifications can miss important patterns, especially in conditions where the course of disease can vary widely between individuals. Traditional methods struggle with messy, real-world data that have irregular timing and many variables. As a result, repeated measures are often underused in research and clinical care. Our project will develop a new AI-based framework called Temporal Pattern Learning (TPL). TPL will use advanced machine learning methods to automatically learn patterns from repeated measures in EHRs, without requiring experts to hand-craft features. With TPL, clinicians could answer questions such as, "Given everything that has happened until today, what is the chance this person's kidney function will drop quickly or that they will need a hospital stay in the next 12 months?". More accurate forecasts mean earlier interventions, better medication choices, and care plans tailored to each individual. TPL has the potential to transform how we use EHR data in both research and care, ultimately improving outcomes for people living with chronic diseases.