Evaluating Patient Representation Strategies for Early Prediction of Alzheimer’s Disease and Related Dementias from Longitudinal EHRs

Alzheimer's disease and related dementias are not sudden events that strike without warning; they are slow, creeping processes that begin changing the brain years before a person ever forgets a name or gets lost in a familiar place. Because these changes happen so early, doctors and researchers have long hoped to find a way to spot the risk long before symptoms appear, allowing for earlier care and better planning. The challenge is that the human brain is complex, and the signs of trouble are often subtle and scattered across a person's medical history. To solve this, scientists have turned to electronic health records, the massive digital archives of every doctor's visit, prescription, and lab test a patient has ever received. These records contain a lifetime of clues, but they are messy, irregular, and difficult to read. The core question for researchers is how to turn this chaotic stream of data into a clear picture of who is at risk, and whether the best way to do that depends on how far in the future we are trying to predict.

A team of researchers set out to test different ways of teaching computers to read these medical histories to predict the onset of dementia. They compared two very different approaches to representing a patient's life in the computer. The first approach was straightforward and transparent: it simply counted how many times a patient had visited the doctor, how many different medications they had taken, and how often they had specific diagnoses, organizing these counts into time windows. The second approach was far more complex, using advanced artificial intelligence models that had been pre-trained on millions of medical records to create a hidden, mathematical "summary" of a patient's entire health journey. The researchers wanted to see which method worked better, how their performance changed as they tried to predict the disease further in advance, and whether these methods would work if applied to a completely different group of patients in a different hospital system.

Using a vast national database of health records, the team trained their models to predict dementia at six, twelve, twenty-four, and thirty-six months before a patient was actually diagnosed. They found that the simpler, count-based approach was generally the most reliable. It was better at distinguishing between patients who would develop the disease and those who would not, and its predictions were more accurate in terms of probability. This advantage was most pronounced when looking at the near future; for predictions made six months before a diagnosis, the simple counting method significantly outperformed the complex artificial intelligence models. However, as the researchers looked further back in time, trying to predict the disease a full three years in advance, the gap between the two methods narrowed. At the thirty-six-month mark, the complex models caught up, performing nearly as well as the simple counts, though they still did not surpass them.

The study also revealed how the computer models "thought" as they looked further back in time. When predicting the disease in the near term, the models relied heavily on specific cognitive symptoms, such as diagnoses of memory loss or mild cognitive impairment. But as the prediction window stretched further into the past, these specific symptoms were often not yet present in the records. In those cases, the models shifted their attention to broader patterns of healthcare use. They began to weigh factors like how frequently a patient visited the hospital, how many different drugs they were prescribed, and the general volume of their medical interactions. It appears that long before the brain shows clear signs of dementia, the body's overall pattern of seeking care changes, and the models learned to use these general signals as early warnings.

The researchers also tested whether these models could be transferred to a different hospital system without retraining them, a scenario known as zero-shot transfer. They applied the models trained on the national database to a separate, independent dataset from a university hospital. The results were stark: performance dropped significantly for both methods. The models struggled to adapt to the different patient population, which was older and had different racial demographics, as well as different patterns of how doctors recorded information. While the complex artificial intelligence models held up slightly better than the simple counts in this difficult transfer scenario, neither approach proved robust enough to work perfectly across different healthcare systems without adjustment. This suggests that while the technology is powerful, it is not yet a universal tool that can be applied anywhere without careful local tuning.

Ultimately, the work demonstrates that there is no single "best" way to represent a patient's medical history for predicting dementia. The choice depends on the timing of the prediction and the specific setting. For spotting the disease in the near future, simple, interpretable counts of medical events remain the gold standard. For looking further ahead, complex artificial intelligence models become more competitive, offering a different balance of accuracy. The findings suggest that the path to early detection lies not in finding a single perfect algorithm, but in understanding how different tools capture different signals at different times, and in recognizing that the signals themselves change as the disease moves from a distant risk to an imminent reality.