DBDP-autonomic Notes

Introduction

• Chronic physical and mental health conditions are increasingly prevalent (1).
• Integrating biological, behavioral, and environmental data is crucial for:
  • Early detection.
  • Just-in-time intervention.
  • Outcome monitoring.
• Mobile monitoring technologies are transforming data collection in non-clinical settings.
• Challenges hindering progress:
  • Misinterpretation of peripheral psychophysiological signals.
  • Lack of transparency and reproducibility in biobehavioral research.

The Criticality of Context in ANS Data Interpretation

• The Autonomic Nervous System (ANS) regulates various physiological processes, including heart rate, blood pressure, respiration, and digestion.
• Its primary function is to maintain homeostasis.
• Understanding the context of ANS data collection is crucial for accurate biobehavioral interpretations.
• Factors influencing ANS activity:
  • Stress, affect, cognition.
  • Physical activity, sleep, illness, medications.
  • Environmental demands.
  • Age, genetics, and health conditions (8, 9).
• ANS responses vary over time and between individuals.
• Misinterpreting transient states can occur when context is excluded (e.g., affect, cognition, physical perturbations) (10, 11, 12).
• Context helps differentiate stimulus-driven shifts from natural fluctuations.
• Including context enables quantitative assessment of interventions.
• Software is needed to enable data fusion and multi-modal analysis.
• Integration of data types could improve digital health initiatives involving ANS data.

Addressing Reproducibility

• Scientific endeavors face a crisis of method and results reproducibility (13, 14).
• Large-scale projects indicate non-replicability in multiple fields (15–19).
• Reported results are frequently incorrect or misstated (20).
• Reasons for irreproducibility:
  • Lack of interoperability between software tools and data sets.
  • Limited record-keeping for complicated data sets.
• Crowdsourced analysis projects reveal analytical flexibility in complex analyses (e.g., 29 teams analyzing identical dataset, with odds ratios for effects ranging from $0:89$ to $2:93$, $M = 1:31$ (21)).
• Reproducibility requires identical statistical analyses.
• The reproducibility crisis is pronounced in ANS data collection and analysis using consumer sensors.
• The field lacks clear guidelines for analyzing autonomic data.
• Researchers repeatedly build new algorithms without comparisons to existing approaches.

Envisioning an Open-Source Data Processing Framework

• Proposed open-source framework to address standardization, interpretation, and reproducibility challenges.
• Identified 10 common problem areas with existing software:
  • Focus on limited biosignals.
  • "Freemium" models.
  • Designed for small lab datasets.
  • No integration across biosignals.
  • No support for open scientific principles.
  • Unsupported, lacking documentation.
  • Inability to analyze context.
  • Static, not designed for community contributions.
  • No ability to archive analysis pathways.
  • Command-line based, challenging for non-programmers.
• Proprietary software has similar issues, plus higher costs and closed code.
• A Survey of User Needs (SUN) confirmed analytical barriers experienced by researchers and engineers.
• Survey demographics : $n = 421$, 31% Researchers, 69% Engineers.
• Survey respondents were from the Society of Psychophysiological Research and the IEEE International Machine Learning for Signal Processing Workshop.
• Over 70% of researchers faced difficulties syncing data, identifying errors, and combining data from different devices.
• Respondents found software tools hard to learn and lacking clear instructions.
• Researchers would benefit from a multimodal data fusion platform with an open-source codebase.
• SUN respondents were enthusiastic about a community-driven open-source framework.

Core Components of the Framework

Community Driven

• Sustainability achieved by inviting scientists to contribute methodologies and algorithms as plugins.
• Researchers can contribute individual tasks or complete end-to-end pipelines.
• Collaboration addresses reproducibility challenges.
• Engineering and computer scientists can validate and refine algorithms.
• Behavioral scientists and clinicians gain access to cutting-edge tools.
• The framework includes a foundational set of validated tools.
• Contributed methods undergo validation through benchmarking datasets and automated testing pipelines.
• Each plugin includes metadata specifying its purpose and validation outcomes.
• A continuous improvement cycle is created through community feedback.
• Standardized templates, documentation, and input/output formats facilitate contributions.
• Data supply chain (38) steps are saved as metadata for transparency and reproducibility.

Data Quality Auditing and Preprocessing

• The framework audits data quality by identifying motion artifacts, environmental factors, and hardware limitations.
• Semi-automated modules for data cleaning, preprocessing, and artifact removal are implemented (39–41).

Signal Segmentation and Alignment

• The framework includes a module for determining appropriate time windows for signal segmentation.
• The module helps users select optimal window lengths for their research needs.
• Tools are included to support improved signal alignment (42).

Contextual Information Integration

• Interpreting physiological data requires understanding the recording context.
• Context includes environmental characteristics and personal factors (affect, physical activity, posture).
• Integrating contextual features improves stress detection (12).
• Plots of physiological data with visual overlays describing recording context are provided.

Data Fusion and Signal Alignment

• This involves aligning data from different sensors or modalities, which might vary in sampling frequencies and timestamps.
• Community-contributed plugins can handle the intricacies of multimodal data by harmonizing signals (short & long term).

Programming Language and GUI

• A graphical user interface (GUI) is needed for open-source physiological processing tools.
• The framework is accessible through both a GUI and a command-line interface (CLI).
• Common open programming languages (Python, R, Bash, etc.) are used.

Science Gateways and Open Science Integration

• The framework aligns with Open Science Framework standards.
• Integration includes leveraging the Digital Health Data Repository.

DBDP Autonomic

• Expanding the Digital Biomarker Discovery Project (DBDP (43)) to include dedicated processing of ANS signals.
• DBDP serves as a hub for collaborative and open research in digital health.
• The code repository includes computational building blocks for common measures of ambulatory physiological data.
  • photoplethysmography (PPG)
  • electrocardiography (ECG)
  • electrodermal activity (EDA).
• The repository comprises four modules:
  • exploratory data analysis
  • data preprocessing
  • feature engineering
  • machine learning model development
• DBDP hosts an archive of code repositories and a list of open-source digital health data.
• DBDP is also developing a code-free GUI-based platform (DBDP Discovery).
• Uploaded code adheres to the framework’s plugin guidelines, allowing it to be cloned as a submodule.
• DBDP-Autonomic specifically targets signals from the autonomic nervous system (ANS) to derive insights into psychological states.

The Role of Context in DBDP-Autonomic

• A key challenge in analyzing ANS data is the role of context.
• DBDP-Autonomic addresses this by integrating contextual data (e.g., activity type, location, and environmental factors).
• DBDP-Autonomic emphasizes the combination of multi-dimensional signals to create unified constructs that reflect psychological states more comprehensively.

Discussion and Call to Community Action

• DBDP has established itself as an open-source hub for digital health.
• DBDP Autonomic could address challenges associated with ANS signal analysis.
• Engagement from the community is vital to the long-term viability of DBDP Autonomic.
• Envisioned member engagement follows the Center for Scientific Collaboration and Community Engagement (CSCCE) Community Participation Model (44).
• Members can:
  • CONVEY/CONSUME: Engage with educational resources.
  • CONTRIBUTE: Add data cleaning algorithms, feature selection methods, and machine learning models.
  • COLLABORATE: Undertake joint research initiatives.
  • CO-CREATE: Organize and lead workshops and working groups.

Conclusion

• A collaborative effort of the DBDP Autonomic community could enable more robust, transparent, and reproducible research in biobehavioral health.
• By emphasizing collaboration, transparency, and rigor, this resource could improve our understanding of complex biobehavioral health issues.
• Such a framework is essential for unlocking the full potential of mobile devices to benefit individual and community health.