Abstract
Adding the PredictND tool to the clinical evaluation increased clinicians' confidence. Furthermore, the results indicate that the tool has the potential to improve prediction of progression for patients with more certain classifications.
Introduction
Predicting whether early-stage cognitive impairment will progress to dementia is a significant challenge in clinical practice. Patients presenting with Subjective Cognitive Decline (SCD) or Mild Cognitive Impairment (MCI) often face an uncertain future; while these conditions can be precursors to Alzheimer’s Disease, many patients remain stable or never develop dementia. Clinicians currently navigate a "data overload" of demographics, neuropsychological tests, MRI scans, and cerebrospinal fluid (CSF) biomarkers without a gold-standard method for integration. This study validates PredictND, a clinical decision support tool designed to synthesize multi-modal data into a reliable prediction.
Methods
This prospective multicenter study followed 429 patients (230 SCD, 199 MCI) across four European memory clinics for an average of 1.7 years. The PredictND tool utilizes a Disease State Index (DSI) classifier that compares individual patient data against a database of confirmed cases.
A key output of the tool is the Disease State Fingerprint (DSF), a visual tree structure. The DSF illustrates which biomarkers—such as MRI volumes or CSF levels—support a "stable" (blue) or "progressive" (red) diagnosis.
Clinicians first provided a prognosis and confidence rating (via Visual Analogue Scale) based on standard evaluation. They then repeated the process after reviewing the PredictND output.
Results
The integration of the PredictND tool primarily served as a "second opinion." While overall diagnostic accuracy remained stable (76% to 77%), the tool significantly impacted specific subgroups and clinician confidence.
Performance in High-Certainty Cases
The tool was most effective when DSI values were at extremes (< 0.2 or > 0.8). In this high-certainty subgroup (n=203), prognostic accuracy improved from 90% to 93%. Notably, for these patients, the tool's standalone accuracy outperformed clinicians working without it by 6.4% ($p = 0.004$).
Clinician Confidence and Reassurance
Even when the final prediction did not change, clinician confidence increased significantly across the cohort ($\Delta VAS = 4%, p < .0001$), particularly for SCD patients. The tool demonstrated a high Negative Predictive Value (NPV), making it exceptionally useful for identifying stable patients and providing clinical reassurance.
Discussion and Limitations
PredictND functions effectively as a validator of clinical judgment. However, the study noted limitations, including a relatively short 1.7-year follow-up period and potential "carry-over" bias in clinician evaluations. Future work will focus on optimizing DSI cut-off values specifically for SCD versus MCI populations to account for varying baseline risks. Despite these hurdles, the tool provides a vital data-backed boost to clinical decision-making and patient communication.
Summary updated