Detecting neurodegenerative disorders from web search signals

Ryen W. White, P. Murali Doraiswamy, Eric HorvitzNpj Digit. Medicine 2018 · DOI · OpenReview

Introduction

Neurodegenerative disorders (NDs), such as Parkinson’s Disease (PD) and Alzheimer’s Disease (AD), present a significant diagnostic challenge. Symptoms often develop insidiously, frequently being dismissed as normal aging until substantial neurological damage has occurred. Current clinical assessments are often "snapshots" in time, which may fail to capture the subtle, longitudinal changes in motor and cognitive function that characterize early-stage NDs. This research explores the potential of "digital phenotypes"—the continuous, observational data generated through daily interactions with technology—as a non-invasive screening tool for these conditions.

Methodology

Researchers from Microsoft and Duke University conducted a large-scale study using 18 months of de-identified search activity from Microsoft Bing, involving over 31.3 million users. To establish a ground truth, the team identified 703 users who issued "experiential diagnostic queries" (e.g., "just diagnosed with Parkinson's"), filtering out those searching on behalf of others.

The study employed Multiple Additive Regression Trees (MART) to classify users based on four distinct feature categories extracted from search logs:

  1. Motor: Fine-grained interaction data including cursor speed, scrolling patterns, and horizontal/vertical oscillations (tremors).
  2. Repetition: Cognitive markers such as repeating the same query or clicking the same result multiple times within a session.
  3. Symptom: The presence of specific query terms related to ND symptoms.
  4. Risk Factors: Inferred demographics and mentions of known risks like head trauma.
Top-ranked features used in the classification model, categorized by type and weighted by importance.

Key Results and Insights

The primary classifier for Parkinson’s Disease achieved an AUROC of 0.9357, demonstrating high sensitivity even at low false-positive rates. Preliminary results for Alzheimer’s Disease were similarly robust, with an AUROC of 0.9135.

The Mouse as a Medical Sensor

A pivotal finding was the efficacy of standard mouse cursor logs in capturing physiological tremors. Motor features were among the most discriminative in the model. This suggests that common hardware can serve as a proxy for clinical motor examinations.

Distinguishing "Look-alikes"

The model proved highly effective at distinguishing PD from Essential Tremor (ET), a condition frequently misdiagnosed as Parkinson's in primary care. The discrimination task between ET and PD achieved an AUROC of 0.9205.

ROC curves for the ET versus PD discrimination task, illustrating the performance impact of various feature sets.

Feature Importance and Ablation

Ablation studies highlighted the critical role of motor and repetition data. When motor features were removed from the PD vs. ET task, performance dropped by approximately 23% (as seen in the shift from the "All minus Repetition" curve to the "All minus Repetition and Motor" curve in the ROC analysis). This confirms that physical interaction patterns are often more informative than the semantic content of search queries.

Discussion and Limitations

While these results are promising, several hurdles remain before clinical implementation. The study relied on self-reported diagnostic queries as a proxy for clinical gold standards, which introduces noise. Furthermore, the massive class imbalance (703 positive cases vs. 31 million negative) requires sophisticated sampling to ensure real-world reliability.

Privacy and ethics remain paramount. The authors propose this technology not as a surveillance system, but as an opt-in, adjunctive screening tool. Future work will involve prospective studies with clinically diagnosed patients to further validate these digital signals against formal medical records.

Summary updated

Abstract

Neurodegenerative disorders, such as Parkinson’s disease (PD) and Alzheimer’s disease (AD), are important public health problems warranting early detection. We trained machine-learned classifiers on the longitudinal search logs of 31,321,773 search engine users to automatically detect neurodegenerative disorders. Several digital phenotypes with high discriminatory weights for detecting these disorders are identified. Classifier sensitivities for PD detection are 94.2/83.1/42.0/34.6% at false positive rates (FPRs) of 20/10/1/0.1%, respectively. Preliminary analysis shows similar performance for AD detection. Subject to further refinement of accuracy and reproducibility, these findings show the promise of web search digital phenotypes as adjunctive screening tools for neurodegenerative disorders.
BibTeX
@article{white2018detecting,
  author = {Ryen W. White and P. Murali Doraiswamy and Eric Horvitz},
  title = {{Detecting neurodegenerative disorders from web search signals}},
  journal = {Npj Digit. Medicine},
  year = {2018},
  doi = {10.1038/s41746-018-0016-6},
}
Figure