Approach to building a Data Warehouse and tackling patient non-compliance

Data Source Layer

• Healthcare data, primarily PHI (protected health information) and non-PHI data were sourced from several decentralized and diverse sources such as patient portals, mammograms, radiologist reports, clinical notes etc.

Data Integration and ETL

• As a first step, we at Finarb have done Data Modeling. Data types that needed to be captured have been identified, and a dimensional data model comprising several dimension tables, such as root-level PHI data and high-level data related to the frequency of mammography screenings, have been built.
• Then, we have extracted and ingested structured and unstructured healthcare data from disparate source databases, maintaining stringent data security measures.
• After that, we have ensured control over healthcare data loading/management.
• Maintaining Data Integrity – We have conducted stringent data validation checks, resolving data anomalies and inconsistencies and addressing data quality issues.

Data Security and Regulatory Compliance

• We have built the DW on HIPAA-compliant Azure Data Factory and have hosted it on secure Azure Cloud.
• For safeguarding data privacy, Finarb has established robust guidelines for protecting PHI data assets, such as PHI and non-PHI raw data encryption methods, at rest and before being transferred to Azure Data Factory.
• We have carried out data transformation in compliance with healthcare regulations (HIPAA, FDA, Hl7 requirements).
• The final DW has been hosted inside the client’s private VNet, giving complete control to the client over their data.

Data storage & Scalability

• Finarb has stored the data on a centralized structured Azure SQL database in a ready-to-be-analyzed form. The DW has been made scalable to handle increased data volume and is hosted on Azure Cloud.
• Future vital integrations to this DW have been made possible through microservices architecture and APIs.

Feature engineering and predictive modeling for risk stratification of patients

• We have chosen the most important variables or features directly or indirectly impacting a patient's compliance.
• Key identified features include actual screening procedure time, patient wait time, delay in procedure start, appointment timing, weekday/weekend examinations, number of patients handled by staff before the current patient, etc.
• Finarb's predictive model has been trained using the feature-engineered dataset to accurately predict patients' likelihood of compliance. This has enabled identifying high-risk patients who were more likely to be non-compliant.

Identification of Compliance Drivers & Recommended Interventions

• We have further clustered high-risk patients into sub-segments based on their top drivers of non-compliance. For each such sub-segment, we have trained the model to find common denominator driver(s), which could be used to customize interventions for improving patient compliance.