Artificial Intelligence in Neurology: Interpreting 2025 Studies in Alzheimer’s Disease, Parkinson’s Disease and Acute Ischaemic Stroke

The most useful question in neurological artificial intelligence (AI) is not whether a model can produce an impressive metric, but whether it can add dependable information at the point of care. Three 2025 studies offer relevant early evidence in Alzheimer’s disease, Parkinson’s disease and acute ischaemic stroke, while also showing why validation and implementation matter as much as algorithmic performance.

Why this matters

AI in neurology is often presented through broad claims about earlier diagnosis or personalised care. The supplied evidence is more specific, and more useful when read that way. One study classifies labelled MRI images across four Alzheimer’s disease-related categories. Another analyses gait signals to identify Parkinson’s disease and assign Hoehn and Yahr (H&Y) severity. A third estimates tissue and discharge clinical outcomes after endovascular thrombectomy (EVT) in acute ischaemic stroke. These are not interchangeable applications. They differ in the data they require, the clinical uncertainty they address and the actionability of any resulting output.

For specialists, the attraction is understandable. MRI contains high-dimensional anatomical information. Wearable sensors and instrumented walkways can quantify aspects of gait that may otherwise be assessed more subjectively. In large vessel occlusion stroke, imaging already informs urgent decisions, while benefit from reperfusion can vary substantially between patients. AI may eventually help organise or quantify information from these sources. Yet the reports reviewed here do not establish that a model result improves diagnostic assessment, disease monitoring, thrombectomy selection or patient outcomes.

The route from model development to practice is particularly demanding in neurology. MRI protocols may vary across services, and sensor availability determines whether gait-derived data can be captured consistently. Acute stroke pathways already rely on rapid imaging, transfer and treatment decisions; any prediction must align with established timelines and should not introduce delay. Performance may not transfer cleanly between datasets with different scanner vendors, field strengths, imaging protocols and patient mixes, or between gait systems with different sensors and collection conditions. Existing imaging interpretation and specialist decision-making remain fundamental. Retrospective performance is necessary evidence, but it is not sufficient evidence for adoption.

Evidence at a glance

Study-by-study clinical interpretation

Alruily et al.: an MRI ensemble for Alzheimer’s disease categories

Alruily and colleagues developed an ensemble model using 6,400 MRI images from a Kaggle dataset. The images were assigned to non-demented (ND), very mild demented (VMD), mild demented (MD) and moderately demented (MOD) categories. In one sentence, the technical approach combined three CNN architectures—VGG16, MobileNet and InceptionResNetV2—pre-trained on ImageNet, with their feature outputs concatenated and fine-tuned for four-class classification.

The headline result was high internal performance: one reported evaluation gave 97.93% accuracy, 98.04% specificity and 95.89% sensitivity. The source also reports a 95.39% multi-classification success rate and a separate evaluation with other performance values. These figures show that the model performed well within the supplied dataset and testing arrangements. They should not be interpreted as a directly established accuracy for clinical diagnosis of Alzheimer’s disease in a new patient population.

The dataset itself is central to that distinction. Of the original images, 3,200 were ND and 64 were MOD; VMD and MD accounted for 2,240 and 896 images, respectively. This is a pronounced imbalance, particularly in the MOD group. The investigators used the Synthetic Minority Oversampling Technique (SMOTE) to create a more balanced distribution. While such balancing can reduce the dominance of majority categories in model development, it does not add independent patient examples to a minority category. The source describes separate evaluation arrangements: an 80%/20% train-test split, an 80%/10%/10% training, testing and validation arrangement, and a balanced test-set account after oversampling. These descriptions reinforce the need to avoid over-reading a single summary figure.

The ensemble’s complexity is also clinically relevant. VGG16 is described as detailed but computationally heavy; MobileNet is more efficient but potentially less able to capture complex information; and InceptionResNetV2 can represent complex image patterns but requires considerable computational resources. Combining the three architectures may enhance feature coverage, but also increases deployment complexity. The model is described as a black box. Grad-CAM and related explainable AI approaches are discussed as suggested interpretability methods, not as demonstrated explanations of model decisions.

Clinical interpretation: this study alone shows that three ImageNet-pre-trained CNN architectures can be combined to classify four labels in a retrospective, single-source Kaggle MRI dataset with high reported internal performance. It does not show that the model has independent patient-level validity, reduces misdiagnosis, improves early clinical detection, or should replace specialist assessment and standard diagnostic processes.

Parkinson’s disease: gait signals as structured clinical data

The Parkinson’s disease investigation took a different route, analysing vGRF signals during walking. Dataset 1 from PhysioNet comprised 93 people with Parkinson’s disease and 73 healthy controls. Dataset 2 from Figshare comprised 112 people with Parkinson’s disease and 110 healthy controls. The study compared decision tree, random forest, XGBoost and Light Gradient Boosting Machine (LightGBM) classifiers for Parkinson’s disease detection and H&Y severity classification.

LightGBM produced the highest reported performance, with 98.25% accuracy in Dataset 1 and 85% in Dataset 2. This difference is arguably the study’s most practical lesson. It demonstrates that strong results in one cohort do not guarantee identical results in another, even when both datasets use vGRF signals. The source attributes the concern broadly to dataset-specific factors, including differences in populations, sensor types and collection protocols. It also reports generally low standard deviations across validation scenarios. Such stability is useful within the tested setting, but cannot substitute for evidence of performance across services and devices.

The study incorporated explainable AI methods and reported that principal components of gait parameters were most associated with disease severity. Stride-time variability, stance-time asymmetry and double-support duration were identified among relevant features. This feature-based interpretation may assist clinical understanding of what the model is using, and it aligns with the study’s focus on known gait abnormalities. However, it does not mean that a gait output represents the complete clinical state of a person with Parkinson’s disease.

The potential contribution is as an objective, quantitative complement to clinical motor assessment. The boundaries are equally clear. The model focused on gait, excluding non-motor manifestations such as cognitive impairment, depression and sleep disturbance, and it did not assess all motor features, including tremor, rigidity and upper-limb bradykinesia. The cross-sectional design does not demonstrate monitoring of progression or response to treatment. Specialised sensors or walkways may also limit accessibility.

von Braun et al.: predicting thrombectomy-related outcomes

Von Braun and colleagues assessed deep learning for outcome prediction in 405 patients with acute proximal anterior-circulation vessel occlusion treated with mechanical thrombectomy. The framework integrated non-contrast CT, CT angiography and CT perfusion (computed tomography perfusion (CTP)) maps, including cerebral blood flow, cerebral blood volume and time-to-maximum. The imaging branches were processed and fused within a three-dimensional CNN.

The key innovation was conditional prediction. The model estimated final infarct volume and NIHSS score at discharge under assumed successful and unsuccessful recanalisation. The difference between those outputs was intended to quantify expected treatment benefit. This is a meaningful research question in EVT, where outcome can vary widely. It is not, however, evidence that clinicians should use the model to overrule existing selection or treatment processes.

For final infarct prediction, mean Dice scores were 0.48 in the internal test set and 0.52 in the external test set, reflecting moderate spatial overlap between predicted and observed infarct regions. The source reports better topographical accuracy than simple thresholding of perfusion maps. For discharge NIHSS prediction, the median absolute error was 1.5–3.0 points. The study reported that perfusion parameter maps were sufficient for tissue prediction and that clinical and demographic variables, along with non-contrast CT and angiography source images, were not required for that specific task.

Routine multimodal CT acquisition makes this work relevant to established stroke imaging. Yet its clinical role remains unproven. The endpoint was NIHSS at discharge, not longer-term functional outcome. The study was retrospective, the data were mainly from a single or limited number of centres, and real-time integration, decision impact and patient outcomes were not evaluated. As with the Alzheimer’s disease model, limited model transparency remains a barrier to interpretation.

Why the studies should or should not be compared directly

The studies share a broad AI-in-neurology theme but should not be ranked by their numerical results. The Alzheimer’s disease and Parkinson’s disease investigations are classification studies with different labels, cohorts and sources of data. The stroke investigation predicts a spatial tissue outcome and a continuous clinical score after a treatment procedure. Accuracy is appropriate for the first two tasks, whereas Dice similarity and median absolute error address different forms of prediction error in the third.

Their validation settings are also distinct. The Alzheimer’s disease model used a single-source image dataset and oversampling in the context of marked class imbalance. The Parkinson’s disease study used two datasets and observed lower performance in the second. The stroke study had internal and external test sets but had limited centre representation. MRI, gait signals and multimodal CT have different acquisition dependencies, and the clinical consequences of error are not equivalent. The studies can be compared as examples of varied research approaches, but not as common measures of clinical utility or readiness.

What this means for clinical practice

These reports offer a practical message for specialists: AI may become useful where it adds clearly defined information to established evaluation, but the evidence here remains pre-implementation. The Alzheimer’s disease work supports further study of MRI classification models, including independent patient-level validation. The Parkinson’s disease study supports research on gait-derived quantitative measures that may complement, rather than duplicate, clinical assessment. The stroke model supports further exploration of patient-specific imaging-based estimates after EVT.

No report demonstrates that its algorithm improves clinical decisions, shortens pathways, reduces workload in practice or improves outcomes. An eventual clinical tool would need dependable data capture, local performance evaluation, clear output presentation, clinician oversight and compatibility with electronic health records, picture archiving and communication systems and existing imaging workflows. The source mentions these integration needs but does not test them. Specialist assessment, standard imaging interpretation and established acute stroke decision timelines therefore remain the relevant clinical foundation.

What remains uncertain

Local validation is unresolved for all three applications. MRI data may differ by imaging protocol, scanner vendor and field strength; CTP and other acute-stroke imaging may also vary by site; wearable gait data depend on sensor availability and collection methods. Governance of imaging and wearable data, documentation of model inputs and outputs, auditability and accountability require attention before implementation. The supplied studies did not evaluate how these requirements would operate in local services or, in stroke, whether predictions can be returned reliably within urgent care timelines.

Representativeness requires further scrutiny. The source identifies potential concerns related to age, sex, race, ethnicity and comorbidities, but does not report subgroup results sufficient to establish equitable performance, calibration or safeguards for under-represented groups. Prospective validation is needed, with multi-centre assessment of transferability and clinical workflow. Longitudinal work is particularly relevant to proposed Parkinson’s disease monitoring, while long-term functional outcomes remain relevant to the stroke model. These are research needs, not gaps that can be filled by retrospective metrics alone.

Conclusion

AI research in neurology is producing increasingly focused models, but their value depends on the clinical problem and evidence context. The Alzheimer’s disease study reported high internal MRI classification performance from a single, imbalanced dataset; the Parkinson’s disease study demonstrated both strong gait-based detection and meaningful variation between datasets; and the stroke study estimated tissue and discharge outcomes after EVT from multimodal CT. Together, they show technical promise across image, sensor and treatment-response data. They do not establish clinical effectiveness, equity, workflow reliability or replacement of specialist judgement. Independent validation, prospective evaluation, interpretable outputs and careful integration with existing pathways are needed before these research findings can support routine use.

Educational content only. It does not replace clinical assessment, current guidelines, or patient-specific professional advice.