A Framework for Generating Explanations from Temporal Personal Health Data

Jonathan J. Harris, Ching-Hua Chen, Mohammed J. ZakiarXiv 2020 · arXiv

Introduction

In the era of the "quantified self," smartphones and wearables collect vast amounts of personal health data. However, a significant gap exists between data collection and actionable understanding. While most applications provide raw charts or basic trend lines, they often fail to explain the underlying "why" or identify complex behavioral correlations. Existing automated summarization methods typically rely on data-heavy neural models that produce low-quality text or rigid rule-based systems incapable of handling complex time-series data. This paper proposes a middle ground: a rule-based, unsupervised framework that generates high-quality, explainable natural language summaries from multivariate temporal health data.

Methodology: The Data-to-Narrative Pipeline

The framework employs a multi-step pipeline to transform numeric sequences into human-readable narratives.

Symbolic Representation and Pattern Mining

To manage noise and varying scales, the system utilizes Symbolic Aggregate approXimation (SAX) to discretize numeric data into symbolic sequences. This allows for analysis at multiple granularities, such as daily versus weekly patterns. Pattern mining is then conducted using the SPADE algorithm for frequent sequence mining (e.g., identifying "if-then" behavioral patterns) and the Squeezer clustering algorithm to find analogous time windows.

Protoforms and Fuzzy Logic

The core of the linguistic generation lies in protoforms—templates with placeholders such as: "On [quantifier] [days] in the past [week], your [attribute] was [summarizer]." The system uses fuzzy logic to select the most accurate linguistic quantifiers (e.g., "most of the time" vs. "some of the time") based on a calculated "Degree of Truth."

The hierarchy of protoforms used to categorize different types of health data summaries.

Provenance

To ensure transparency and build user trust, every generated summary is linked back to the source data. The system highlights relevant data points on the corresponding charts, providing immediate visual evidence for the textual claims. This approach moves the AI from a "black box" to an "open book."

Evaluation and Results

The framework was evaluated using two large-scale real-world datasets: MyFitnessPal (587,187 days of food logs) and Insight4Wear (11.5 million activity and heart rate records). Generalizability was further tested using weather and stock market data.

Multivariate health data showing calorie and carbohydrate intake over 25 weeks, used as input for the summarization framework.

A user study with 11 participants yielded high scores for readability and comprehensiveness (4.13/5). Notably, "Goal Evaluation" summaries—which assess progress toward specific health targets—were rated most useful (4.55/5). Quantitatively, the rule-based approach proved faster and more reliable in text quality than state-of-the-art neural methods for this specific domain.

Discussion and Future Work

The system excels at multivariate discovery, identifying hidden correlations across different metrics, such as carbohydrate spikes coinciding with high calorie intake on specific days. Furthermore, it supports population-level insights, allowing for comparisons between individual data and broader group trends.

Despite its successes, the framework currently relies on manually designed protoforms. Future research could explore using deep learning to automatically discover these narrative structures. Additionally, while individual summaries are accurate, improving the narrative flow to link these summaries into a cohesive "story" remains a key objective for future development.

Summary updated

BibTeX
@misc{harris2020framework,
  author = {Jonathan J. Harris and Ching-Hua Chen and Mohammed J. Zaki},
  title = {{A Framework for Generating Explanations from Temporal Personal Health Data}},
  year = {2020},
  eprint = {2003.09530},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2003.09530},
}
Figure