Personalized Hypothesis Tests for Detecting Medication Response in Parkinson Disease Patients Using iPhone Sensor Data

Elias Chaibub Neto, Brian M. Bot, Thanneer M. Perumal, Larsson Omberg, Justin Guinney, Mike Kellen, Arno Klein, Stephen H. Friend, Andrew D. TristerOpenReview 2015 · DOI · OpenReview

Introduction

Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by motor symptoms that vary significantly between individuals. While dopaminergic medications like levodopa are standard treatments, their efficacy often wanes over time, leading to complications like dyskinesia. Clinicians face the challenge of determining whether a specific patient is responding optimally to their medication regimen. Traditional population-level studies often fail to capture these highly individual responses.

With the advent of smartphone-based health monitoring, such as Apple’s ResearchKit, it is now possible to collect high-frequency, longitudinal data from patients in their home environments. However, this data is often "messy"—characterized by serial correlations, irregular sampling, and missing labels. This paper introduces "N-of-1" personalized statistical tests designed to detect medication response in individual PD patients using smartphone tapping sensor data.

Methodology

The researchers developed two primary "personalized" testing frameworks that treat each patient as an independent study:

1. Union-Intersection Tests

This approach extracts multiple features from the 20-second tapping tasks (e.g., total tap count, mean inter-tap interval). A standard statistical test (like a t-test or Wilcoxon rank-sum test) is applied to each feature to compare "before medication" and "after medication" states. To maintain a global significance level across multiple features, the Benjamini-Hochberg (BH) correction is applied.

2. Classifier-Based Tests

Instead of testing features individually, this method trains a machine learning classifier (e.g., Random Forest or Extra Trees) to distinguish between the "before" and "after" labels using the entire feature set. The test statistic is the Area Under the ROC Curve (AUROC). A permutation-based approach is used to determine if the observed AUROC is significantly higher than 0.5 (random chance).

Visualization of a single patient's feature time series before and after medication.

Theoretical Foundation: The Role of Exchangeability

A major hurdle in applying standard tests to longitudinal sensor data is the violation of the independence assumption. Measurements taken close together in time are typically correlated. The authors prove that these tests remain valid as long as the data labels are exchangeable under the null hypothesis of "no medication response."

If medication has no effect, the "before" and "after" labels are essentially arbitrary. Shuffling these labels does not change the joint distribution of the data. This insight allows for the use of robust, well-understood statistical tools on complex, serially correlated time-series data without inflating false positive rates.

Comparison of p-value distributions under different dependency structures. Panel (f) shows that when both data and labels share the same dependency (exchangeability), the test remains valid (uniform p-values).

Results and Insights

The methods were validated using both simulated data and real-world data from the mPower study, involving 57 patients.

  • Detection of Responders: Using a conservative significance threshold ($p < 0.001$), the personalized tests identified approximately 25% of the mPower cohort as having a statistically significant response to medication.
  • Power Trade-offs: Simulation studies indicated that Union-Intersection tests are more powerful when only a few features are informative. Conversely, classifier-based tests (particularly tree-based models) excel when the signal is distributed across many features, as they avoid the heavy penalties associated with multiple-testing corrections.
  • Feature Importance: In patients identified as responders, "number of taps" and "mean tapping interval" were the most discriminative features, aligning with clinical expectations of improved motor function post-medication.

Conclusion and Future Work

This research demonstrates that personalized, N-of-1 hypothesis tests are a viable and statistically sound method for monitoring medication response in Parkinson’s disease. By leveraging the principle of exchangeability, clinicians can utilize smartphone sensor data to make data-driven, individualized treatment decisions.

Future work aims to extend these tests to detect the direction of response (improvement vs. worsening) and to integrate multi-modal data from gait and voice sensors for a more comprehensive clinical picture.

Summary updated

Abstract

We propose hypothesis tests for detecting dopaminergic medication response in Parkinson disease patients, using longitudinal sensor data collected by smartphones. The processed data is composed of multiple features extracted from active tapping tasks performed by the participant on a daily basis, before and after medication, over several months. Each extracted feature corresponds to a time series of measurements annotated according to whether the measurement was taken before or after the patient has taken his/her medication. Even though the data is longitudinal in nature, we show that simple hypothesis tests for detecting medication response, which ignore the serial correlation structure of the data, are still statistically valid, showing type I error rates at the nominal level. We propose two distinct personalized testing approaches. In the first, we combine multiple feature-specific tests into a single union-intersection test. In the second, we construct personalized classifiers of the before/after medication labels using all the extracted features of a given participant, and test the null hypothesis that the area under the receiver operating characteristic curve of the classifier is equal to 1/2. We compare the statistical power of the personalized classifier tests and personalized union-intersection tests in a simulation study, and illustrate the performance of the proposed tests using data from mPower Parkinsons disease study, recently launched as part of Apples ResearchKit mobile platform. Our results suggest that the personalized tests, which ignore the longitudinal aspect of the data, can perform well in real data analyses, suggesting they might be used as a sound baseline approach, to which more sophisticated methods can be compared to.
BibTeX
@misc{neto2015personalized,
  author = {Elias Chaibub Neto and Brian M. Bot and Thanneer M. Perumal and Larsson Omberg and Justin Guinney and Mike Kellen and Arno Klein and Stephen H. Friend and Andrew D. Trister},
  title = {{Personalized Hypothesis Tests for Detecting Medication Response in Parkinson Disease Patients Using iPhone Sensor Data}},
  year = {2015},
  doi = {10.1142/9789814749411\_0026},
  howpublished = {OpenReview},
}
Figure