@bmarebwa and I edited a topic collection centered around AI-assisted identification of novel multimodal imaging markers and underlying mechanisms in PD for npj Parkinson’s Disease (AI-assisted identification of novel multimodal imaging markers and underlying mechanisms in PD) a little over two years ago, the goal was modest on paper(s) and ambitious in practice; bring together work showing how machine learning is changing what we can see, measure, and predict in PD neuroimaging. More than 35 submitted manuscripts to this topic, and twelve accepted articles later; a few threads have emerged that are worth pulling out explicitly, because taken together they say more than any single paper does on its own.
Several papers used AI to extract signal that wasn’t previously accessible, rather than just classifying existing scans. Shin et al.'s deep learning–based susceptibility source separation for in-vivo iron mapping, Xiong et al.'s 7T structural MRI work (Support vector machine-driven Parkinson’s disease identification: a 7-Tesla multidimensional structural MRI approach | npj Parkinson's Disease), and two transcranial sonography papers (Kang et al.'s multi-classifier fusion, Zhao et al.'s super-resolution grading system) all push toward more quantitative, reproducible measurement, expanding the measurable phenotype rather than just diagnosing from it.
The papers by Shi et al. (brain-first vs. body-first asymmetry), Wang et al. (EEG + gait cross-attention subtyping), and Zhang et al. (tremor network differentiation of PD vs. essential tremor) all reflect a shift from “does this person have PD” toward “which biological process is driving it.” That’s a more tractable near-term application than the disease-modifying endpoints everyone eventually wants, especially in light of recent shift towards a more biological definition of the disease.
In Hähnel et al.'s brain age gap as a progression marker based on PPMI data (Brain age gap as predictor of disease progression in Parkinson’s disease | npj Parkinson's Disease), and the NCER-PD consortium’s MCI decision-support tool (Martínez Tirado et al.) show models built for longitudinal utility. Leavitt et al.'s DBS targeting optimization is the clearest example of AI feeding directly into treatment rather than diagnosis, and Barber-Janer et al.'s CNN-based mouse histopathology work is a very good example that the imaging-to-pathology bridge deserves more attention than it usually gets.
However, external validation is still a key issue. Most models here are trained and tested within a single cohort; PPMI-scale, prospectively harmonized, cross-cohort testing remains the exception. Interpretability is inconsistent, and performance against atypical parkinsonism look-alikes is rarely reported at the scale needed to change practice rather than opinion.
One major takeaway from this journey is that AI’s most durable contribution so far isn’t a diagnostic classifier, rather it’s expanding what counts as a measurable feature in the first place: iron separated from myelin, tremor networks separated from each other, subtype separated from stage. The diagnostic payoff follows from that, not the other way around.
Check out the collection and let us know your thoughts 
This is a really fascinating conclusion derived from your unique perspective from working on this special issue. Thanks for highlighting it for us! I have earmarked a number of those papers to dig into in more detail.
In the meantime, I will be thinking more about the potential for AI/ML to uncover useful features instead of focusing on diagnostic outcomes. This makes intuitive sense to me, as we are still just uncovering the various ways of defining and staging Parkinson’s. There is not (yet) one “killer feature” which clearly defines those with PD, stages, and outcomes, so perhaps we need to revisit and optimize the features further before we can make more accurate diagnostic tools.
Personally, I will keep this in mind as we define the next stages of development work for the PIE and PIE-clean repositories, as they handle both ML and data cleaning/feature engineering respectively!