ClinicLens: Visual Analytics for Exploring and Optimizing the Testing Capacity of Clinics given Uncertainty

Yu Dong, Jie Liang, Longbing Cao, Daniel R. CatchpoolearXiv 2023 · DOI · arXiv

Introduction

During the COVID-19 pandemic, testing clinics were the primary defense against viral spread. However, optimizing their capacity proved challenging for healthcare administrators due to significant data uncertainty. Key hurdles included attribution gaps (data reported at regional rather than clinic levels), reporting delays, and the complex interplay of operational features like business hours and government interventions. ClinicLens addresses these challenges by providing a visual analytics system that models and optimizes clinic-specific features to meet future demand through "what-if" simulations.

System Architecture and Methodology

ClinicLens utilizes a dual-component architecture to bridge the gap between raw data and actionable policy insights.

The ClinicLens system architecture, featuring a back-end engine for data processing and a front-end visualization suite.

Back-end Engine

The back-end handles data ingestion, cleaning, and feature engineering. To resolve the "multiple-clinics-to-one-LGA" attribution problem, the system employs a feature modeling component that rescales individual clinic features into aggregate Local Government Area (LGA) datasets for model training. The system supports various regression models, with Random Forest and XGBoost selected for their high predictive accuracy and ability to rank feature importance.

Front-end Visualization

The visual analytics interface consists of three primary views:

  1. Map-Lens: A geospatial interface providing nested layers for intervention timelines, test/case counts, and positive rates.
  2. Indented Tree-Matrix: A hierarchical visualization for comparing clinic features such as accessibility and booking requirements across different regions.
  3. Testing Capacities Prediction View: An interactive simulation environment where users can adjust clinic parameters and observe predicted impacts on regional testing throughput compared to historical ground truth.

Datasets and Feature Engineering

The system was validated using a comprehensive dataset from New South Wales (NSW), Australia, spanning January 2020 to October 2022. This included:

  • Government COVID-19 Data: Daily testing and case counts.
  • Clinic Service Data: Operational details for 248 clinics, including business hours and service types.
  • Demographic Data: 2021 Census information.
  • Intervention Data: Unstructured data extracted via Natural Language Processing (NLP) from news sources to categorize restriction levels.

Results and Model Performance

Quantitative evaluation demonstrated that ensemble learning methods are highly effective for predicting testing capacity despite data noise. Random Forest emerged as the top-performing model, achieving an $R^2$ of 0.99 and the lowest Root Mean Square Error (RMSE).

Performance comparison of various regression models used in ClinicLens.

Qualitative validation with domain experts—including data scientists and clinic managers—confirmed the system's utility. Experts noted that the "mobile lens" and "indented tree-matrix" revealed operational bottlenecks, such as the disproportionate impact of "testing anchor" clinics, which were previously obscured in traditional data formats.

Insights and Future Work

A key finding from the ClinicLens simulations is the "ripple effect" of operational changes: adjusting business hours on a single day influences predicted test counts over a subsequent 7-day period due to reporting lags and human behavior. Furthermore, while extending weekend hours significantly boosts capacity, the model suggests that short mid-day breaks have a negligible negative impact on total throughput.

Future development will focus on real-time data integration across other regions and the implementation of a "clinic editor" to simulate the impact of opening new facilities or modifying more granular operational variables.

Summary updated

BibTeX
@misc{dong2023cliniclens,
  author = {Yu Dong and Jie Liang and Longbing Cao and Daniel R. Catchpoole},
  title = {{ClinicLens: Visual Analytics for Exploring and Optimizing the Testing Capacity of Clinics given Uncertainty}},
  year = {2023},
  eprint = {2303.13558},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2303.13558},
}
Figure