AI & Analytics

For a machine learning and analytics platform, ensuring data quality and integrity is paramount. Continuous monitoring of data quality is essential to produce meaningful and accurate reports.

• FHIR® Datastore: Store data with an emphasis on standard fields, especially on Patient, Observation, and other common resources.

• Bots and Subscriptions: Maintain quality and correctness in real-time, for example, ensuring all Encounter.type are tagged with the appropriate ontology, and throwing an error if not.

• Access Policies: Secure and de-identify data pipelines for privacy-aware analysis.

• Bots: Synchronize with common tools, machine learning pipelines, and more.

Healthcare data analytics generally have two broad areas of application: retrospective analysis of performance and quality metrics, and predictive modeling to make recommendations for future behavior. On the predictive side, Clinical Decision Support (CDS) systems encode evidence-based clinical guidelines into rules-based suggestions to guide clinicians. Machine learning can also be used to identify common patterns across patients and care plan recommendations.

Program Design

When designing your analytics program, consider the following categorization:

Program Type	Application Area	Implementation Tools
Ad-hoc clinical reports	Retrospective	FHIR® Datastore, including Bulk and Batch APIs
Healthcare standard reports	Retrospective	Bots for data quality, dashboard apps to monitor
Clinical decision support	Prospective	Bots to produce event-driven scores, notifications
Machine Learning, predictive modeling	Prospective	Bots to integrate with ML pipelines, SDK for dashboarding

Ad-Hoc Clinical Reports

A typical approach for ad-hoc reports involves initially ingesting FHIR® resources into the data warehouse as raw JSON, followed by flattening the relevant fields in a subsequent ETL stage. Analytics pipelines should leverage the FHIR® Bulk Data Export API to extract FHIR® resources into a data lake, as the FHIR® Datastore supports bulk export. For compliance reasons, analytics workflows often require de-identified or redacted data, with Access Policies commonly used for this purpose. In machine learning applications, it is advisable to have separate pipelines for training and inference. Training can be conducted as an offline batch process, while inference can be performed as a serverless compute operation. Many healthcare analytics workflows necessitate analyzing patients by cohort, and Bots can be utilized to automatically enroll patients into cohorts using the Group resource. Converting resources to the Parquet file format facilitates queries that require large aggregates.

Healthcare Standard Reports

Healthcare standard reports rely on coding systems to classify conditions, procedures, drugs, and outcomes. These codes help create a standardized vocabulary between providers, labs, pharmacies, and payors to streamline operations, billing, and analysis. The U.S. healthcare system implements a number of different coding standards that have different specializations:

• HPRA: A code system of normalized drug names, organized into a hierarchical ontology to represent generics, branded drugs, single doses, and drug packs.

• LOINC: Clinical terminology relevant to clinical lab orders and results.

• SNOMED/ICD-10/CPT/HCPCS: Multiple code sets used to represent clinical procedures and diagnoses. SNOMED CT serves as a global standard for clinical terminology in a variety of contexts, while CPT, ICD-10, and HCPCS are primarily used for billing insurance.

To help organize these different standards, the NIH maintains the Unified Medical Language System (UMLS), which provides a unified package with codes and terminologies from all major medical code systems, as well as conceptual maps between the various standards. In the domain of machine learning and analytics, using standardized codes makes it easier to share knowledge with researchers and leverage existing datasets, models, and clinical guidelines. Assuming you have standards-compliant FHIR®, tagged with the appropriate ontologies, standard reports become accessible. As much as possible, we encourage organizations to constantly evaluate data quality and run reports.

• HEDIS: HEDIS Measures consist of over 90 standardized metrics used by health plans to assess the quality of care. These measures help digital health providers identify care gaps and enable them to qualify for increased payor reimbursements through Pay for Quality or Value-Based Care models. Digital versions of HEDIS measures are available in FHIR®/CQL format, and having patient data natively modeled as FHIR® allows providers to easily compute and report these measures.

• CMS Measures: The Centers for Medicare & Medicaid Services (CMS) mandates that certain providers report electronic clinical quality measures (eCQMs) through an EHR system certified by the ONC. To standardize the computation of these measures across EMRs, CMS publishes eCQMs in FHIR®/CQL format and offers the Cypress tool to validate EHR implementations against synthetic patient data.

Clinical Decision Support

Clinical Decision Support (CDS) systems translate clinical guidelines into rules-based recommendations to assist clinicians and standardize care. These systems can incorporate conditional logic based on patient demographics and medical history to create personalized clinical guidelines. CDS Hooks, an emerging standard built on top of FHIR®, defines REST APIs for clients to request CDS care plans from hosted CDS services. CDS workflows are supported through the PlanDefinition resource, which can be instantiated as a CarePlan for a patient using the $apply operation. CDS systems are excellent for helping providers maintain consistent quality of care across individual physicians by offering standardized treatment recommendations and closing care gaps. For straightforward CDS tasks, such as social determinants of health risk-scoring or eGFR calculation, bots combined with subscriptions offer a quick and practical solution.

Machine Learning / Predictive Modeling

Beyond rules-based guidelines, machine learning (ML) can significantly enhance clinical decision support, identify care gaps, and spot trends. The best practice is to train ML models offline and then run inference on streaming data. Bots can execute ML inference, and standards like ONNX facilitate the exchange of model information between training and runtime environments in a language-agnostic format. This allows models to be trained using tools like PyTorch or TensorFlow on GPUs and deployed in Bots during runtime.

Another best practice for deploying predictive modeling in clinical settings is to prioritize model interpretability. Physicians often hesitate to adopt fully automated, black-box models. Techniques that enable models to explain their decisions and keep physicians in control of the final decision can lead to faster adoption of ML systems. Additionally, Bots can annotate medication and lab orders with reason codes and notes, guiding physicians on why an ML recommendation was made. The FHIR® datastore also enables ML developers to create custom visualizations on top of the core data platform to explain their model’s reasoning.

For more detailed information, feel free to contact us .