From algorithms to agreements: Why life sciences must rethink the evidence framework
May 2026 | Written by Michael Grabner, PhD
Life sciences leaders have moved past debating whether artificial intelligence (AI), real-world evidence (RWE), or value-based healthcare (VBHC) matters. The question now is how quickly these capabilities can be embedded across development, regulatory, and commercialization strategies to remain competitive. ranked AI #1, RWE #2, and VBHC #3.¹ This marks a structural shift in how therapeutic value is generated, demonstrated, and sustained in the market.
Traditional ways of proving value, such as relying on clinical trials and fixed evidence reports, are insufficient on their own. Regulators, payers, and providers increasingly expect continuous, real-world validation of clinical and economic outcomes. The future value in life sciences depends on bringing together predictive algorithms, high-quality real-world data (RWD), and outcomes-based agreements into a single, cohesive evidence strategy. Predictive analytics can improve trial design and patient selection, while robust RWD enables ongoing evidence generation beyond clinical trials. These aren’t separate efforts. Together, these capabilities support performance-based agreements that link reimbursement to how therapies perform in real-world populations. They create a more relevant and sophisticated way to show value, helping life sciences companies move beyond one-time evidence submissions to continuously demonstrate how their products perform over time.
AI is not the strategy; deployment is
AI’s ascent to the top of ISPOR’s trends list reflects its expanding footprint across systematic reviews, predictive modeling, trial optimization, and analytics acceleration.1 Yet, many of the industry's conversations remain model-centric. A shift is underway from model development to operational deployment, where value is realized only when algorithms are validated, scalable, and embedded in real-world decision-making.
Large-scale validation analyses illustrate this transition. In a recent Carelon Research study, the Klinrisk machine learning model for chronic kidney disease (CKD) progression achieved strong performance across more than 4 million individuals spanning commercial, Medicare, and Medicaid populations, with AUCs ranging from 0.80 to 0.87.2 Its importance extends beyond predictive performance. By relying on routinely collected laboratory and claims data and demonstrating consistency across diverse populations, it shows what deployment-ready AI looks like in practice.
This shift introduces new requirements. First, external validation is essential. Models must prove reliability across real-world populations before they can be trusted in clinical or payer settings.2 Second, scalability depends on data accessibility. Models built on widely available inputs are far more deployable than those requiring specialized or sparse data. Third, integration into workflows is critical. Machine learning creates value only when predictions inform action, such as identifying high-risk patients early enough to enable intervention.
Another example concerns obesity identification. Administrative claims alone often underreport obesity, limiting population health management. Machine learning models that integrate claims, clinical data, laboratory results, and social drivers of health (SDOH) significantly improve identification and outperform the use of diagnosis codes alone.3 These approaches demonstrate how AI can correct structural data gaps when designed to operate within RWD limitations.
For payers and manufacturers, the implication is clear. The question is whether AI’s risk predictions can be implemented, trusted, and used at scale. Competitive advantages will come not from building better models in isolation, but from deploying them effectively within care management, evidence generation, and value-based decision-making.
RWD at the core of modern evidence generation
RWD has evolved into a comprehensive, research-ready asset that can capture experiences and outcomes relevant to patients, providers, and payers, enabling the generation of actionable RWE. Our Healthcare Integrated Research Database (HIRD®) illustrates what mature RWD infrastructure looks like today.4 As of Q1 2026, the HIRD includes more than 94 million individuals currently and formerly enrolled in commercial and managed Medicare health plans located throughout the U.S., including 24 million actively enrolled members. Demographic distributions closely mirror U.S. Census benchmarks for age, sex, and region, enhancing generalizability.4 However, scale alone is not the defining feature. Three attributes matter more for payer and pharma decision-makers: continuity, traceability, and integration.
Long-term, continuous enrollment enhances the ability to evaluate long-term health outcomes with greater confidence, including outcomes not typically captured in clinical trials. Median continuous enrollment among actively enrolled members in the HIRD exceeds three years, with 36% enrolled for ≥ 5 years.5 Longer follow-up can enhance treatment persistence analyses, safety monitoring, and total cost-of-care modeling.
Traceability and auditability are other factors. Data in the HIRD can be traced back to their sources, including closed claims, enrollment files, electronic health records (EHR), publicly available data, and others. Monthly quality control procedures monitor completeness and accuracy.4 In one example, discrepancies in migraine dosing patterns were traced to free sample use, improving interpretation of safety findings.6 In an era of regulatory scrutiny, the ability to trace data back to its source is essential.
Lastly, multi-modal integration is necessary. The HIRD integrates claims with structured and unstructured EHR data, laboratory results, mortality data validated against the National Death Index, oncology staging information, and SDOH.4 Beyond these existing data linkages, its infrastructure supports the integration of other data, including patient-reported outcomes (PROs) via direct research engagement with key stakeholders. Because the HIRD contains individually identifiable information, it enables recruitment of members and providers into prospective studies, with the ability to link study findings back to historical claims data. This combination of multi-source data and embedded research capabilities transforms administrative datasets into comprehensive, patient-level ecosystems that support continuous evidence generation.
Carelon Research and the HIRD contribute to the FDA Sentinel, the Biologics Effectiveness and Safety system, IMEDS and NEST initiatives.4 That involvement signals a broader reality: RWE is integral to regulatory science. For payers negotiating contracts and manufacturers defending value propositions, regulatory-grade RWE platforms are essential rather than supplementary assets.
VBHC is becoming contract design
Today, VBHC is increasingly operationalized through contract design. While aligning payment to outcomes is familiar, the real shift is the ability to structure and refine value-based agreements (VBAs) using RWD.
A collaboration between Elevance Health (CarelonRx and Carelon Research) and Takeda illustrates this evolution. What began in 2017 as a continuation-of-treatment agreement that shared risk when patients discontinued therapy early has progressed into a multi-phase, data-driven contract framework.7
The initial phase focused on a measurable and operational outcome: treatment persistence. By using clearly defined metrics and existing claims data, the agreement balanced simplicity with credibility, enabling accurate adjudication and building trust between partners.7
As RWE accumulated, the agreement evolved. In the next phase, a predictive model estimated expected disease-related medical costs using claims-based risk stratification. When actual costs were lower than these expectations, the savings were shared between payer and manufacturer. This marked a shift from drug-specific guarantees to accountability for broader cost outcomes, supported by independent validation and transparent analytics.7
The most advanced iteration expanded the definition of value further by incorporating additional clinically relevant cost drivers into a more comprehensive total cost-of-care model. Importantly, this did not come at the expense of operational feasibility, highlighting the need to balance precision with practicality.7
Across all phases, the agreement functioned as a continuous learning system. RWD informed adjudication, generated insights, and enabled ongoing contract redesign. This progression allowed stakeholders to build confidence, align value definitions, and adapt as evidence evolved.7
VBHC has evolved from intent to execution. It requires predictive analytics, longitudinal data, and collaborative design. In this model, value is not just measured; it is contractually defined, tested, and continuously refined.
The integrated evidence framework
AI helps identify at-risk populations and can uncover data blind spots. Actionable RWE provides continuous, verifiable measurement. VBHC aligns financial incentives around outcomes. HEOR integrates these layers, ensuring methodological rigor and interpretability.
Healthcare is transitioning from retrospective analysis to predictive intelligence. From siloed data assets to integrated evidence structures. Life science companies and healthcare payers that embrace this alignment will shape the next generation of healthcare values. The industry’s task is not merely to adopt AI, build databases, or sign performance-based agreements. It is to integrate them into a coherent strategy. Value is now measurable, auditable, and increasingly contractable. The organizations that master the full evidence package will define the future of healthcare.
Acknowledgments
We thank Sarah Daugherty, PhD, MPH for her valued input on patient-centered research.
Sources
- ISPOR. ISPOR 2026–2027 Top 10 HEOR Trends. International Society for Pharmacoeconomics and Outcomes Research (Apr 16, 2026): https://www.ispor.org/heor-resources/top-10-heor-trends.
- Tangri, N., Ferguson, T. W., Teng, C.-C., Bamforth, R. J., Smith, J. L., Guzman, M., & Goss, A. Validation of the Klinrisk machine learning model for CKD progression in a large representative US population. Journal of the American Society of Nephrology (Aug 2025): https://pubmed.ncbi.nlm.nih.gov/40768294.
- Head, M. A., Chhibber, A., Venkataraman, M., Kuti, E., Teng, C.-C., Lartey, L., Tan, H., & Donato, B. M. K. Real-world identification of obesity: Development of a claims-based machine learning algorithm. Poster number 378. Presented at ObesityWeek 2025. Atlanta, GA, United States.
- Barron, J. J., Willey, V. J., Doherty, B. T., Tunceli, O., Waltz, C. R., Grabner, M., Beachler, D. C., Lanes, S., & Cziraky, M. J. The Healthcare Integrated Research Database (HIRD) as a real-world data source for pharmacoepidemiologic research. Pharmacoepidemiology and Drug Safety (Jan 2025): https://pubmed.ncbi.nlm.nih.gov/39909722.
- Barron JJ, Venkataraman M, Grabner M, Harris KM, Willey V, Beachler DC. Does health plan enrollment data support long-term follow-up of patients with chronic disease? Assessment of the Healthcare Integrated Research Database (HIRD®). Poster number RWD118. Presented at: ISPOR 2025 (May 2025); Montreal, QC, Canada.
- Hoffman SR, Lanes S, Crowe CL, Daniels K, Mawanda F, Schroeder KM. Distribution of samples of a new preventive migraine drug and potential implications for validity of pharmacoepidemiology studies using administrative data. Poster number 170. Presented at: International Society for Pharmacoepidemiology (ISPE) Annual Conference (2022).
- Thomas N, Sefain H, Stanek E, et al. Real-world insights in action: a model program for executing successful and sustainable value-based agreements. Managed Healthcare Executive (Feb 2026).