Digital Data and a path toward a Healthcare Framework

Douglas Drake, Digital Strategy Lead, NMS Health 

Some Quotes:

• 30% of the world´s data volume is being generated by healthcare insights. 
• Humans are expected to reach more than 1,400 digital device interactions per person per day by the end of 2020, and nearly 5000 in just five more years. 
• People are living longer; populations in emerging markets are exponential; drug development costs are exploding. 
• Global healthcare budget will be an estimated USD 15 trillion in 2030
• Healthcare technology will, one day very soon, literally touch us all daily. 

*https://www.rbccm.com/en/gib/healthcare/episode/the_healthcare_data_explosion

Background

Increasingly, understanding the patient’s journey through their healthcare system is being recognized as critical to modeling and understanding different patient outcomes and care metrics. Patients are complex, with not just a single factor between diagnosis, treatment, and outcome. Especially for chronic diseases, the social demographics, the pharmacogenomics, the time between diagnosis and treatment, the distance to a healthcare provider, existing comorbidities and health risks play in the longitudinal metrics of care and probable outcomes.

Healthcare has become multifactorial and is by nature an open system, with dynamic, sometimes unpredictable, and open chaotic behavior. The recent and ongoing COVID-19 pandemic is a perfect example of this: the evolution and spread of the vaccine has not been predictable or anticipated, nor has the uptake of vaccine treatments been global or sufficient on a voluntary basis to prevent the further spread of the virus.

A doctor’s first assessment of a patient often refers to any previous records possible with the first review looking for any immediate changes or reported conditions. However, this retrospective review must now become more detailed and more extensive, especially considering multiple conditions or comorbidities the patient may present. This is often more than a single physician in a single visit can manage but is increasingly available through digital data sources (provided below) that physicians, caregivers, epidemiologists, and healthcare researchers can use to model, test, and validate various healthcare questions.

• pharmacy claims data analyses 

• electronic medical record database analyses 

• retrospective patient chart reviews 

• analysis of registry data 

• cohorts or longitudinal studies

Joyeux A, Olivaris R, Understand the care pathway/patient journey, available at https://rwe-navigator.eu/using-the-navigator-decision-support-tool/clarify-the-issues/understanding-the-patient-journey/

Smart Hospitals and Economics 

Hospitals and physician clinics are often responsible for managing and maintaining patient healthcare data but are not designed as IT services, but for directed patient care. A review by McKinsey and Co in 2019 (see figure below), highlights the need for digital connectivity within hospitals, in a model called smart hospitals, but also highlights the changing economics in healthcare. As patient care becomes more decentralized, care services are moving out of traditional hospital care to other retail services.  

This poses a challenge to established hospitals and hospital services that need to be technology innovative as at the same time reimbursement costs are being challenged and there is increased competition to deliver outpatient healthcare. 

One model to compensate and offset potential lost revenue by developing clinical research services that can often bring external funding into traditional bed settings. This market is expected to reach USD 52 Billion by 2025* and largely driven by the need to find specific patients for clinical trial research in specific markets to support late-stage medication development, regulatory approval, and price negotiation. 

*https://www.globenewswire.com/en/news-release/2023/07/07/2701067/0/en/Clinical-Trials-Market-is-Expected-to-Reach-52-0-billion-MarketsandMarkets.html

As patient journeys have become more individualized, treatment has become more personalized, so have medication approaches and ultimately medication development strategies. However, finding specific patients in a timely manner that are readily available for specific enrollment deadlines is critically challenging, much like finding needles in not just one haystack, but multiple and scattered hay rolls across Switzerland.  This dictates the hospitals that want to work with external research sponsors, have a framework that will allow them to quickly review and qualify their expertise and their patients for sponsored studies. Ultimately, this also brings value back to the hospital and care system by better enabling retrospective review, understanding their own care metrics and better also targeted outreach to at risk patients for care follow up.  

*Finding the future of care provision: The role of smart hospitals, available at https://www.mckinsey.com/industries/healthcare/our-insights/finding-the-future-of-care-provision-the-role-of-smart-hospitals#/

The remainder of this paper will focus on the elements of a common data structure and suggestions necessary for hospitals to enable external partner research capability and patient enablement. 

The intent is to highlight current trends, capabilities and features that will be important for hospital and care giver consideration. 

Hospital Information Systems

Hospital information systems (HIS), used to capture patient record detail, are notoriously large, disparate, and often decentralized. It is not uncommon for different services within a Healthcare system to have separate systems depending upon their therapeutic expertise and data capture needs. HIS´s are designed for capturing patient healthcare data but not specifically for ready querying and data visibility. As a single patient may encounter very different capabilities within a care system, from in patient, or as outpatient, in emergency care, or by referral, blood enzyme levels by laboratory testing, biopsy screens, procedures such as imaging or x-ray, physiotherapy, or chemotherapy, it is not uncommon for a patient´s records to be contained across different systems holding different types of healthcare data. 

As the patient journey is obviously a composite of all these elements, the goal is to be able to aggregate these various data sources into a single query-able source that can be used by the hospital and enable perspective for outside partnership. As the hospital and their staff and the patient caregivers, those with care privileges may have permission to access and view patient records for retrospective as well as prospective treatment decision. However, patient healthcare records should not leave the hospital IT space and firewall as that could reveal personal and private patient details in violation of GDPR law as well as country and local data protections. 

A common data architecture to enable this is shown below, in which the HIS or data warehouse (DWH) is connected to a Query Server (in blue), still within the hospital firewall (in gray), via a directional extraction and transformational layer (ETL in orange) that processes patient records into anonymized healthcare records as a nightly service. The Query Server is accessible external to hospital firewall, only via a secure private cloud service with no inbound port capability. This allows external queries to be downloaded and only query results to be uploaded without any primary data extraction. 

Data Types and Ontology 

The reimbursable aspects of healthcare have been coded over now over 30 years as various electronic medical record coding structures, which include patient demographics, diagnosis codes, prescribed medications, procedures, and laboratory tests. These digital records, originally intended for insurance claims for care reimbursement, are a tremendous recourse of data that often have secondary use to map patient journey, care patterns as well as potential interventional care. The figure below highlights the various dimensions of EMR data which are generated as part of a patient's record over time and can give insights to patient journey, standard of care and outcome metrics. 

The Observational Medical Outcomes Partnership (OMOP) has developed the Common Data Model (CDM) and standardized vocabularies for medical terms used across various clinical domains. See https://www.ohdsi.org/data-standardization/ and for open-source tools support. OMOP is supported by the European Health Data & Evidence Network (EHDEN) which has also provided grants to hospitals in the past to index their data using OMOP to create common ontology datasets for global RWD analytics. See: https://www.ehden.eu/consortium-partners/

Physician Charts

EMR codes provide direct metrics on patient care, but do not provide insight into physician interpretation of the patient such as risk factors, social demographics, stage of disease, range of motion, mobility, family history, interpretation of medical images, biopsies, genetic screens, progression or stability of the patient, response to treatment or potential adverse events. These metrics are captured often in the physician´s chart or patient notes that include the summary reports of various specialty testing, imaging, biopsy, or interpretive diagnostics. Increasingly these insights are critical to understanding and specifying patient cohorts, stratifying care and both retrospective research as well as prospective clinical research matching. 

Physician charts often contain unstructured data elements as the files, even if electronic, are formatted as free text combined with medical terms in relational sentence structures, which are linguistically structured, but not in a standard ontology or relationship mapping. There have been strong advances in large language models (LLMs), ChatGTP as an example, to understand and interpret language but these need to be applied to healthcare data in a structured fashion and in other languages, not exclusively English. However, in combination with other terminology mappings, such as Human Phenotype Ontology (HPO), see https://hpo.jax.org/app/ may be able to accelerate the mapping of specific disease trait descriptions to disease diagnosis. The eventual goal is to use these technologies together in specific Natural Language Processing (NLP) tools to extract key interpretative metrics from text documents that can supplement the EMR coding with specific interpretative detail. 

Medical Images and specialty reports

Medical images and specialty reports such as pathology, biopsy, genetic screening need to be archived and accessible, but a summary level data extract needs to be either highlighted or generated that can be included in an NLP data ETL effort. The data architecture model is to keep these primary images and reports as archival material using the summary level generation as feed into the structured and relational data structure with EMR. There is increasing interest in accessing medical images, especially in oncology, to confirm interpretation and disease stage and progression, as well as develop AI methodology toward better image-based diagnostics, so linking is important, but not direct incorporation to a relational schema. 

Data Architecture

Due to the sheer and disparate amount of data within healthcare, creating a centralized capability requires a complex architecture with minimally a 3 tier: source, middle tier connectivity/ analytic services and data access layer for specific applications and users. Below is an example, designed to show various scenarios, from ingestion, data storage, analytics and management, publishing, and data access.

Conclusions

The intent within this paper is to discuss some of the needs for hospitals to create data infrastructure that better enable better patient visibility and insight into patients that can contribute to collaborative clinical research. Various data structures and formats have been highlighted as well as workflow using NLP toward enriching existing structured EMR information from care reimbursement. Finally, a high-level architecture has been briefly introduced for the purpose of discussion and various option mappings. The overall intent for this paper is to be a discussion piece and possible framework for the next discussions and a possible workshop with the various stakeholders and contributors.

Acknowledgements and Thanks: Le Vin Chin Chris S. Yves Neuhaus