Predicting Decline Before It Shows Up on the Scan
I first ran into Julian Epstein's work back around 2018 when a colleague forwarded me a paper on motor-cognitive coupling in prodromal Parkinson's. At the time, the predictive models he was building felt borderline absurd. We were talking about detecting imminent cognitive decline months before standard clinical assessment would flag anything. Most of us in the field weren't buying it until we started testing it ourselves. The core of Epstein's approach sits at the intersection of longitudinal neuropsychological assessment and machine learning classification. He didn't invent the algorithms himself. What he did was build one of the most meticulously annotated longitudinal datasets in behavioral neurology and then layer predictive modeling on top of it. The dataset comes primarily from his work at the University of Pittsburgh Alzheimer's Disease Research Center and related cohorts. Thousands of patient visits, standardized neuropsychological batteries, imaging markers, and eventually survival outcomes, all tied together. What makes his models stand out isn't the sophistication of the algorithm. Random forests and gradient boosters can do that much. It's the feature engineering. Epstein's team treats cognitive test trajectories as continuous signals rather than static snapshots. They look at intra-individual change rates across repeated assessments and feed those into survival models. The models predict things like conversion from mild cognitive impairment to Alzheimer's dementia, time-to-motor-symptom onset in Parkinson's, and hospitalization risk in neurodegenerative populations.
I spent about three weeks trying to reproduce a subset of these models using publicly available data from the ADNI dataset. The process was harder than it should have been. The original paper didn't disclose the exact preprocessing pipeline. Missing data imputation, for instance, was handled differently than standard practice. I tried multiple imputation by chained equations and then k-nearest neighbors, and the model performance swung by nearly fifteen percentage points depending on which method I used. That's a huge variance. I ended up settling on hot-deck imputation constrained by age and education strata, which got me within five percent of the reported AUC. It still wasn't a clean replicate. Here's the thing most people writing about Epstein's work gloss over: these models are highly sensitive to cohort composition. If you validate them on a population with different education levels or comorbidity profiles, the predictions degrade fast. I saw this firsthand when I tried applying a converted model to a rural clinic population in western Pennsylvania. The baseline cognitive scores were lower across the board due to limited healthcare access, not pathology. The model flagged nearly forty percent of patients as high-risk for rapid decline, and about half of those calls turned out to be false positives once we followed them for two years. The fix wasn't architectural. I recalibrated the intercept and added region-of-care as a covariate. That brought the false positive rate down to something usable, maybe twelve to fourteen percent. The features that actually matter in these models are not the ones you'd expect. Everyone assumes MMSE or MoCA scores drive the predictions. They don't, not primarily. Epstein's work consistently shows that reaction time variability, verbal fluency switching costs, and gait-speed-to-cognition ratios carry more discriminative weight than standard memory indices. This is counter-intuitive for clinicians who spend most of their assessment time on memory testing. But when you run a variable importance analysis on these models, the cognitive-motor intersection features dominate. That's been consistent across at least four separate publications from his group.
There's also a limitation worth stating bluntly. These models don't generalize well to early-stage screening in community settings. They're built for clinical research populations where you have repeated measures and controlled assessment protocols. Throw them into a primary care environment where you're doing a single-point screening, and the predictions fall apart. The models need trajectory data. Without at least two baseline assessments spaced six months apart, you're essentially guessing. I learned that the hard way when a startup contacted me about deploying one of these models in their telehealth platform. They only had cross-sectional data. I told them it wouldn't work. They deployed it anyway. Six months later, the false positive rate was around sixty percent, and they pulled the model from production. If you're looking to use Epstein's methodology, the most practical route is starting with the published feature sets rather than trying to reverse-engineer the full pipeline. His group has been relatively transparent about which neuropsychological instruments feed into the models. The Repeatable Battery for the Assessment of Neuropsychological Status, the Digit Symbol Coding subtest from the WAIS, the Verbal Fluency tests, and gait assessment protocols are the backbone. You don't need the exact algorithm. You need the right input features and enough longitudinal data points to calculate change scores. I also want to flag one common mistake I see repeatedly. People try to substitute proxy measures when they can't access the full battery. They use phone-based cognitive tests or app-derived reaction times in place of lab-administered measures. The models were trained on clinician-administered data with standardized conditions. Phone-based tests introduce variance that the model interprets as signal. It creates noise that looks like predictive pattern but isn't. I caught this in a retrospective audit of our own clinic data. About eight percent of our flagged patients turned out to be artifacts of a different testing modality. Once we restricted to in-person assessments, the signal became much cleaner.
Get the Full Details
The downloadable resources are limited. Epstein hasn't released open-source implementations of his models. The closest thing you'll find are the supplementary materials from his papers, which typically include the feature lists and sometimes the hyperparameter settings. For actual model weights or code, your best path is to contact his lab directly at the University of Pittsburgh. They've been reasonable about research collaborations in my experience. I've corresponded with them twice, and they shared preprocessing notes on request, though they didn't provide raw model code. One final note on interpreting the results. The models output probability estimates, not diagnoses. I've seen clinicians and even some researchers treat a high probability score as equivalent to a positive diagnostic finding. That's incorrect and potentially harmful. These models estimate risk trajectories, not disease presence. The difference matters, especially when you're discussing results with patients or using them to guide treatment decisions. Epstein himself has been careful about this framing in his publications, but the gap between what the model does and how it gets interpreted in practice is wider than it should be.
