Understanding Healthcare Data: Stakeholders, Datasets, Bias and AI Opportunities

In the rapidly evolving landscape of healthcare, data plays a pivotal role in improving patient outcomes, enhancing operational efficiency, and driving research. However, navigating this complex ecosystem requires an understanding of the various stakeholders involved, the diversity of datasets generated, the opportunities for artificial intelligence (AI) modelling, and the potential biases that can arise. In this blog, we explore these aspects and discuss strategies to mitigate biases in healthcare data.

Let's start by understanding the complexity of the ecosystem of healthcare.

Key Stakeholders in the Healthcare System

1. Patients: Individuals seeking care, who may decide to pursue treatment based on their health needs or information gathered from various sources, including the internet.
2. Healthcare Providers: This group includes physicians, nurses, and allied health professionals who diagnose conditions, prescribe treatments, and document patient interactions in electronic medical records (EMRs).
3. Pharmacies: Entities that dispense medications to patients, often playing a role in patient education regarding drug use.
4. Pharmaceutical Companies: Organisations involved in the research, development, and manufacturing of medications, influencing treatment options available to patients.
5. Drug Distributors: Companies that supply medications to pharmacies, ensuring that drugs are available for patient use.
6. Pharmacy Benefits Management Companies: These organisations manage prescription drug benefits for health plans, negotiating prices and determining coverage. They try to secure lower drug costs for insurers and insurance companies.
7. Medical Device Companies: Manufacturers of medical equipment and devices that are essential for diagnosis and treatment.
8. Government Agencies: Both state and federal entities that regulate healthcare practices, monitor public health, and collect data to inform policy decisions.
9. Medical Researchers: Individuals or groups conducting studies to advance medical knowledge, often publishing findings that can influence clinical practices and healthcare policies.
10. Health Insurance Companies: Organisations that provide coverage for medical expenses, impacting how care is accessed and reimbursed.
11. Public Health Organisations: Entities focused on improving community health through research, education, and policy advocacy.
12. Policy Makers: They utilise healthcare data to inform public health decisions and policy development, aiming to improve healthcare systems and outcomes.

Understanding the diverse interests of these stakeholders is crucial, as their goals can sometimes conflict, leading to tensions in data generation and usage.

Diversity of Datasets

The healthcare system generates a wide array of datasets, each capturing different aspects of patient care and outcomes. Some notable sources include:

• Electronic Medical Records (EMRs): Comprehensive records of patient interactions with healthcare providers, including diagnoses, treatments, and outcomes.
• Claims Data: Information related to billing and insurance claims, providing insights into healthcare utilisation and costs. Typically less detailed in clinical content this provides comprehensive and representative data for large populations, including the elderly, children, the very poor, and nursing home residents, who are often under-represented or excluded from clinical trials.
• Patient-Generated Data: Data collected directly from patients, such as self-reported symptoms and health status, often through online portals or mobile applications.
• Clinical Trials: Rigorous studies designed to evaluate the effectiveness of treatments, generating high-quality data on patient responses and outcomes.
• Public Health Surveys: National datasets like the National Health and Nutritional Examination Survey (NHANES) that provide insights into population health trends.
• Post-Marketing Surveillance and Disease Monitoring Data Sources: Various government and professional society databases that monitor adverse events of drugs and devices after their approval to ensure safety and take action if serious side effects occur.

Each dataset offers unique opportunities for analysis and insights, but no single source provides a complete picture of a patient's healthcare journey. Therefore, leveraging multiple datasets can lead to more reliable conclusions.

However, given the diversity of stakeholders and the resulting data, there is plenty of scope for introducing bias in these datasets that can impact downstream analysis.

Source of Bias in Data and Impact on AI Modelling

• Patient Selection Bias: A study on diabetes management may only include patients who regularly visit their healthcare provider, excluding those who manage their diabetes through lifestyle changes without medical supervision and those who cannot afford healthcare through the selected provider. This bias can lead to an AI model that overestimates the effectiveness of medical interventions, as it does not account for a significant portion of the population with different management strategies and outcomes. If the social disparity in the population is significant, this can exclude entire demographics.
• Clinical Documentation Bias: Healthcare providers may document only severe cases of a condition, neglecting patients with mild symptoms who do not require extensive treatment. The resulting dataset may lack representation of the full spectrum of the condition, leading to AI models that fail to generalise well across different severities, potentially misguiding treatment recommendations.
• Coding Bias: A coder may mistakenly assign a diagnosis code for hypertension when the patient actually has hypotension or omit a relevant comorbidity due to oversight. This can result in inaccurate datasets that misrepresent patient conditions, leading to flawed predictions regarding treatment efficacy or risk stratification.
• Temporal Bias: If a patient's health status is recorded weeks after a visit, the provider may forget critical details, leading to incomplete or inaccurate records. AI models trained on such data may misinterpret the temporal progression of diseases, affecting predictions about disease trajectories and potentially leading to inappropriate clinical decisions.
• Patient Reporting Bias: A patient may underreport alcohol consumption during a health survey due to social stigma or exaggerate medication adherence to please their healthcare provider. This bias can distort the understanding of patient behaviours and health outcomes, resulting in AI models that inaccurately assess the relationship between lifestyle factors and health, ultimately affecting public health recommendations and interventions.
• Claims Reporting Bias: Due to free refills available from clinics or patients stopping treatment earlier than prescribed, claims data may under or over-report actual drug use. This bias can lead to AI models that inaccurately assess medication adherence and its effects on health outcomes. As a result, the models may inaccurately assess the efficacy of treatments, potentially leading to misguided clinical decisions and ineffective public health strategies.

Mitigating Strategies to Overcome Biases

To enhance the reliability of healthcare data and mitigate biases, several strategies can be employed:

1. Data Triangulation: Utilise multiple data sources to cross-validate findings and reduce reliance on a single dataset. This approach can help identify discrepancies and provide a more comprehensive view of patient care 9.
2. Standardised Data Collection: Implement standardised protocols for data collection and documentation to minimise measurement errors and improve consistency across datasets.
3. Awareness and Training: Educate healthcare providers and researchers about potential biases and their implications, fostering a culture of data integrity and accuracy.
4. Statistical Adjustments: Use statistical techniques to adjust for known biases in analyses, ensuring that results are more representative of the broader population.
5. Diverse Representation in Studies: Ensure that clinical trials and studies include diverse populations to capture a wide range of responses and outcomes. This can help create AI models that are more generalizable and applicable to various patient groups.
6. Continuous Monitoring and Validation: Regularly assess AI models for performance and fairness, making adjustments as necessary to account for new data and emerging biases.

Conclusion

The healthcare data landscape is complex, with diverse stakeholders generating a wealth of information. Understanding the types of datasets available and the potential biases that can arise is crucial for leveraging data and exploiting the opportunities for AI modeling, effectively. By implementing strategies to mitigate biases, we can enhance the quality of healthcare data and drive meaningful improvements in patient care and health outcomes.