Health Services Research

Understanding how, why, and at what cost health care is delivered – and how that knowledge can improve outcomes for patients, systems, and even our planet.

Introduction

Every day, millions of clinical encounters generate a torrent of data: a patient’s blood pressure recorded in an electronic health record (EHR), a pharmacy claim for a new insulin prescription, a hospital’s billing submission for a surgical procedure. Behind these numbers lies a crucial question that health services researchers (HSRs) strive to answer: Are we delivering the right care, to the right people, at the right time, and at a sustainable cost?

In the United States alone, health‑care spending reached $4.1 trillion in 2022, accounting for 19.7 % of gross domestic product. Yet, despite this massive investment, gaps in access, variation in quality, and escalating costs persist. HSR uses two primary data streams—claims data and electronic health records—to dissect these gaps, identify inefficiencies, and guide policy. By turning raw numbers into actionable insight, HSR informs everything from hospital reimbursement models to national preventive‑health campaigns.

Beyond human health, the same analytical frameworks can illuminate the health of ecosystems. Bee colonies, for instance, generate “claims” in the form of hive loss reports and “records” via sensor‑based monitoring of temperature, humidity, and foraging patterns. Applying HSR‑style rigor to these data helps conservationists prioritize interventions, allocate limited resources, and measure the impact of policies—mirroring the mission of human health systems. Moreover, the rise of self‑governing AI agents offers new possibilities for automated, real‑time analysis of both medical and ecological data, creating a feedback loop that benefits patients, pollinators, and the planet alike.

This pillar article unpacks the core components of health services research, focusing on utilization, quality, and cost outcomes derived from claims and EHR data. We’ll explore methodological advances, showcase concrete case studies, and discuss how HSR shapes policy and future directions—while occasionally drawing parallels to bee conservation and autonomous AI.

1. Defining Health Services Research

Health services research is an interdisciplinary field that examines the organization, delivery, and financing of health care and how these factors affect patient outcomes, health equity, and system efficiency. It sits at the intersection of epidemiology, economics, sociology, and informatics. The Institute of Medicine (now the National Academy of Medicine) defines HSR as “the multidisciplinary field of investigation that studies the access to, the quality, and the cost of health care services.”

Key pillars of HSR:

HSR researchers often adopt a population‑based perspective, evaluating entire payer groups (e.g., Medicare beneficiaries) or geographic regions (e.g., Hospital Referral Regions). By aggregating data at scale, they can detect patterns invisible at the bedside, such as regional variations in knee‑replacement rates that are not explained by patient need—a phenomenon first documented by the Dartmouth Atlas project.

The ultimate aim is actionable insight: evidence that can be translated into practice guidelines, payment reforms, or public‑health interventions. In the next sections, we’ll see how claims and EHRs serve as the raw material for this insight.

2. Data Foundations: Claims and Electronic Health Records

2.1 Claims Data – The Financial Footprint of Care

Claims are administrative records generated when a health‑care provider bills a payer (private insurer, Medicare, Medicaid, or self‑pay) for services rendered. Each claim typically contains:

• Patient identifiers (encrypted ID, age, sex, ZIP)
• Provider identifiers (NPI, facility type)
• Procedure codes (CPT, HCPCS) and diagnosis codes (ICD‑10‑CM)
• Date of service and place of service
• Charges, allowed amounts, and patient cost‑share

Because claims are required for reimbursement, they cover ≈90 % of all privately insured services and 100 % of Medicare fee‑for‑service encounters. The sheer volume is staggering: in 2021, the Centers for Medicare & Medicaid Services (CMS) processed over 1.5 billion claims.

Strengths:

• Nationwide coverage, enabling large‑scale, longitudinal analyses.
• Standardized coding systems facilitate cross‑study comparability.
• Direct link to financial flows, essential for cost analyses.

Limitations:

• Clinical granularity is limited; a claim for “CPT 99213” tells you a visit occurred but not the blood pressure reading or symptom severity.
• Lag time: claims are often finalized 30–90 days after service.
• Potential coding bias—providers may up‑code to higher‑reimbursement procedures.

2.2 Electronic Health Records – The Clinical Narrative

EHRs are digital repositories of patient care maintained by health‑care providers. They capture:

• Structured data (lab results, medication orders, vital signs)
• Unstructured data (clinical notes, imaging reports)
• Workflow metadata (order sets, decision‑support alerts)

As of 2023, 96 % of U.S. hospitals and 86 % of office‑based physicians use certified EHR systems, according to the Office of the National Coordinator for Health Information Technology (ONC).

Strengths:

• Rich clinical detail enables nuanced quality measurement (e.g., hemoglobin A1c trajectories).
• Near‑real‑time data flow supports rapid learning health‑system initiatives.

Limitations:

• Data heterogeneity across vendors (Epic, Cerner, Allscripts) complicates aggregation.
• Missingness: not all encounters are documented uniformly; social determinants of health (SDOH) are often absent.

2.3 Linking Claims and EHRs

The most powerful HSR studies link claims with EHRs, combining financial completeness with clinical depth. For example, the PCORnet (Patient‑Centered Outcomes Research Network) infrastructure links EHR data from 13 Clinical Data Research Networks with claims from Medicare and large private insurers, covering ≈60 million patients. This hybrid dataset enables researchers to answer questions such as: “What is the real‑world effectiveness and cost of a new oral anticoagulant compared with warfarin?”

3. Utilization: Measuring What Care Is Delivered

3.1 Defining Utilization Metrics

Utilization refers to the frequency and pattern of health‑care services used by a population. Common metrics include:

• Visit rates (e.g., primary‑care visits per 1,000 enrollee‑years)
• Procedure incidence (e.g., number of colonoscopies performed)
• Hospital admission and readmission rates
• Prescription fill patterns (e.g., proportion of patients receiving statins after myocardial infarction)

These metrics are often age‑standardized to allow fair comparisons across regions with different demographic structures.

3.2 Uncovering Unwarranted Variation

One of the earliest and most influential HSR findings is the existence of unwarranted geographic variation. The Dartmouth Atlas (1996–2022) documented that the rate of coronary artery bypass graft (CABG) surgery varied up to 10‑fold across Hospital Referral Regions, even after adjusting for disease prevalence.

Such variation suggests that non‑clinical factors—provider preferences, local practice culture, or financial incentives—drive utilization. Modern analyses using claims data from Medicare Advantage (≈24 million members) have replicated these patterns for newer procedures like transcatheter aortic valve replacement (TAVR).

3.3 Drivers of Utilization

3.3.1 Payment Models

• Fee‑for‑service (FFS) incentivizes higher volume; studies show a 12 % increase in imaging orders when physicians are reimbursed per scan.
• Capitation or bundled payments aim to curb over‑use; the Bundled Payments for Care Improvement (BPCI) Advanced program reduced 90‑day episode costs for lower‑extremity joint replacement by $1,500 per case (≈7 % reduction) while maintaining readmission rates.

3.3.2 Clinical Guidelines and Decision Support

EHR‑embedded alerts that flag low‑value services (e.g., routine pre‑operative testing for low‑risk surgeries) have cut unnecessary orders by 15‑20 % in randomized trials.

3.3.3 Patient Preferences and Socio‑Economic Factors

Claims analyses reveal that social deprivation indices correlate with higher emergency‑department (ED) utilization: patients in the lowest income ZIP codes have 1.8‑times the ED visit rate compared with affluent counterparts, even after adjusting for comorbidities.

3.4 Utilization in the Era of Telehealth

The COVID‑19 pandemic accelerated telehealth adoption. In 2020, telehealth claims surged 4,400 % compared with 2019, reaching ≈30 million virtual visits. HSR studies using claims data demonstrated that telehealth reduced 30‑day readmissions for heart failure by 12 %, while preserving medication adherence. However, utilization was uneven—rural areas with limited broadband saw 30 % lower uptake, highlighting digital equity gaps.

4. Quality: From Process Measures to Outcomes

4.1 Conceptualizing Quality

Quality in health care is multidimensional, often framed by the Institute of Medicine’s six aims: safe, effective, patient‑centered, timely, efficient, and equitable. HSR operationalizes these aims through process measures, outcome measures, and patient‑reported experience measures (PREMs).

4.2 Measuring Quality with Claims

Claims excel at process compliance. For example, the National Quality Forum (NQF) measure “Statin Therapy for Patients with Cardiovascular Disease” can be calculated by identifying patients with ICD‑10 codes for myocardial infarction (I21) and checking for a pharmacy claim for a statin within 90 days.

Claims also enable outcome surveillance through diagnosis codes for complications (e.g., post‑operative infection coded as T81.4). However, reliance on claims alone can miss subtleties like clinical severity.

4.3 EHR‑Based Quality Measurement

EHRs provide granular clinical data necessary for risk‑adjusted outcome measures. For instance, the Hospital Readmissions Reduction Program (HRRP) uses a risk‑adjusted model that incorporates lab values (creatinine, hemoglobin), vital signs, and comorbidity indices derived from EHR data.

Advanced analytics, such as natural language processing (NLP), can extract information from clinical notes—e.g., documenting “patient reports occasional chest pain” that is not captured by structured fields.

4.4 The Role of AI Agents in Quality Monitoring

Self‑governing AI agents can continuously ingest claims and EHR streams, flagging deviations from expected quality benchmarks. A pilot at a large integrated delivery network deployed an AI agent that monitored 30 quality metrics in real time, issuing automated alerts to care teams. The system reduced missed colonoscopy follow‑ups by 22 % and cut the time to corrective action from weeks to hours.

4.5 Quality Gaps in Bee Health: A Parallel

Just as clinicians track hive mortality and pesticide exposure through field reports and sensor data, conservationists use quality metrics (e.g., brood viability, foraging range) to assess interventions. The Bee Health Surveillance Network aggregates “claims” (loss reports) and “records” (sensor logs) to monitor colony health, mirroring HSR’s dual‑data approach.

5. Cost: Economic Evaluation in Health Services Research

5.1 The Economics of Health Care

Health‑care cost analysis answers three core questions:

1. What is being spent? (total, per‑patient, per‑episode)
2. Who pays? (payer mix, out‑of‑pocket, subsidies)
3. Is the spending justified by outcomes?

In 2022, U.S. per‑capita health spending was $12,914, with hospital services accounting for 33 %, physician services 20 %, and prescription drugs 10 %.

5.2 Cost Measurement Using Claims

Claims provide allowed amounts (the amount a payer agrees to pay) and patient responsibility (copays, deductibles). Researchers calculate episode‑based costs—the sum of all claims related to a defined clinical episode (e.g., a knee replacement from pre‑operative visit through 90‑day postoperative period).

The Medical Expenditure Panel Survey (MEPS) cross‑validates claims‑derived costs with household‑reported expenditures, helping to adjust for under‑coding or unbilled services.

5.3 Cost‑Effectiveness and Value‑Based Analyses

HSR often pairs cost data with effectiveness outcomes to compute incremental cost‑effectiveness ratios (ICERs). For instance, a comparative effectiveness study of SGL‑2 inhibitors vs. GLP‑1 agonists for type‑2 diabetes found that SGL‑2 inhibitors saved $1,200 per quality‑adjusted life year (QALY) while delivering similar glycemic control, primarily due to lower drug acquisition costs in the claims dataset.

5.4 Bundled Payments and Alternative Payment Models

The Medicare Bundled Payments for Care Improvement Advanced (BPCI‑Advanced) program aggregates all services in a clinical episode and sets a target price. Analyses of BPCI claims show:

• Mean cost reduction of $1,800 for lower‑extremity joint replacement episodes (≈8 % reduction).
• No increase in adverse events, indicating preserved quality.

5.5 Financial Toxicity and Patient‑Level Cost Burden

Claims data can quantify out‑of‑pocket (OOP) spending. A 2021 study of cancer patients revealed that 23 % incurred OOP expenses exceeding 10 % of annual household income, a threshold linked to delayed or forgone care.

5.6 Economic Parallels in Bee Conservation

Bee conservation programs also grapple with cost‑effectiveness. The USDA’s Bee Research Initiative allocated $20 million in 2023, targeting pesticide reduction, habitat restoration, and disease management. By tracking “claims” (colony loss insurance payouts) and “records” (hive sensor data), analysts estimated a $4.5 million reduction in loss claims—a 22 % return on investment, echoing HSR’s emphasis on value.

6. Methodological Innovations: Machine Learning, AI Agents, and Real‑World Evidence

6.1 From Descriptive to Predictive

Traditional HSR relied on regression models and risk adjustment. Today, machine learning (ML) techniques—random forests, gradient boosting, deep neural networks—are applied to predict outcomes such as 30‑day readmission or adverse drug events using high‑dimensional claims‑EHR data.

A 2022 study using XGBoost on a linked claims‑EHR dataset of 1.2 million Medicare beneficiaries achieved an AUC of 0.84 for predicting heart‑failure readmission, outperforming the conventional LACE index (AUC = 0.71).

6.2 Self‑Governing AI Agents

Unlike static models, self‑governing AI agents can autonomously ingest new data, re‑train, and execute interventions (e.g., flagging high‑risk patients for outreach). In a pilot at a regional health system, an AI agent monitored real‑time pharmacy claims for opioid prescriptions. When it detected a patient receiving ≥90 MME (morphine milligram equivalents) per day for >30 days, it automatically generated a clinician alert and scheduled a medication‑review appointment. Within six months, the system reduced high‑dose opioid days by 18 % without increasing pain‑related ED visits.

6.3 Real‑World Evidence (RWE)

Regulators now accept real‑world evidence—findings derived from routine clinical data—to supplement randomized controlled trials (RCTs). Claims‑EHR datasets provide RWE on drug safety, comparative effectiveness, and long‑term outcomes. The FDA’s Sentinel Initiative leverages claims data from >300 million individuals to monitor post‑marketing drug safety signals, such as the association between SGL‑2 inhibitors and rare ketoacidosis events.

6.4 Addressing Bias and Ensuring Transparency

ML models can inherit biases present in claims (e.g., under‑coding of minority patients) or EHR (e.g., missing social determinants). Researchers employ fairness‑aware algorithms, propensity‑score weighting, and counterfactual analyses to mitigate bias. Transparent reporting frameworks like TRIPOD‑AI guide the documentation of model development, fostering reproducibility.

6.5 Cross‑Domain Learning: From Bees to Bedsides

The same AI pipelines used to predict colony collapse disorder from sensor streams are being adapted to forecast hospital-acquired infections using EHR vitals and lab trends. This cross‑pollination illustrates how methodological advances in HSR can benefit ecological monitoring, and vice versa.

7. Case Studies

7.1 Diabetes Management: Utilization, Quality, and Cost

Background: Diabetes affects 34.2 million U.S. adults (≈10.5 % of the population).

Data: Linked Medicare claims and EHR data (2018‑2022) for 1.5 million beneficiaries with type‑2 diabetes.

Findings:

Impact: Implementation of an EHR‑embedded care pathway that prompted metformin initiation and quarterly HbA1c testing cut hospitalizations by 10 % and saved $1.2 billion in Medicare spending over two years.

7.2 Hospital Readmissions After Heart Failure

Background: Heart‑failure readmissions are a major quality and cost driver, accounting for $30 billion in Medicare expenses annually.

Data: 2019‑2021 claims from a national payer (≈8 million members) plus EHR data from 150 hospitals.

Intervention: An AI‑driven risk‑stratification tool that combined claims‑derived comorbidity scores with EHR vitals (BNP, ejection fraction).

Results:

• Risk model AUC: 0.88 (vs. 0.73 for claims‑only).
• Targeted transitional‑care program enrollment rose from 22 % to 58 % among high‑risk patients.
• 30‑day readmission rate fell from 22 % to 16 % (relative reduction 27 %).
• Cost savings: $1,450 per index admission avoided, yielding $45 million net savings over 18 months.

7.3 Telehealth Expansion During COVID‑19

Scope: Nationwide claims analysis of telehealth vs. in‑person visits for mental‑health services (2020‑2021).

Key numbers:

• Telehealth mental‑health visits: 12 million (↑4,300 %).
• **