Artificial Intelligence for Precision Allocation of Cardiometabolic Therapies in Diabetes and Metabolic Disease: From Risk Prediction to Treatment Response

Technical Summary: Artificial Intelligence for Precision Allocation of Cardiometabolic Therapies

Problem Statement
The therapeutic landscape for type 2 diabetes, obesity, chronic kidney disease (CKD), heart failure, and metabolic dysfunction-associated steatotic liver disease (MASLD) has shifted from a glucose-centered paradigm to multi-organ risk modification. The expansion of therapies such as GLP-1 receptor agonists (GLP-1RA), SGLT2 inhibitors (SGLT2i), and dual incretin-based agents (e.g., tirzepatide) has created a complex decision-making environment. While these drugs offer benefits in weight reduction, cardiorenal protection, and hepatic risk reduction, clinical benefits remain heterogeneous across patients. Traditional evidence-based medicine, relying on average treatment effects from randomized trials, fails to address individual variations in absolute event risk, metabolic responsiveness, organ-specific benefit, tolerability, and patient preference. The core challenge is determining not if a therapy works, but which patient should receive which therapy, when, and for which dominant clinical priority.

Methodological Framework
The paper proposes a structured operational framework for AI-driven precision allocation, distinguishing four distinct analytic questions that must be separated before integration:

1. Risk Prediction: Identifying who is likely to develop complications under usual care (untreated risk).
2. Treatment Prioritization: Determining who requires earlier or more intensive intervention due to high untreated risk.
3. Post-Treatment Outcome Prediction: Estimating who is likely to achieve a favorable outcome after exposure to a specific therapy.
4. Causal Treatment-Effect Estimation: Estimating the conditional average treatment effect (CATE)—the expected difference in outcome under one strategy versus another for a specific patient.

The authors emphasize that precision allocation requires linking these layers without conflating prognosis with drug-specific benefit. The framework includes a "benefit-arbitration" step that integrates safety filters (contraindications, organ thresholds), expected causal benefits, and competing harms (adverse effects, cost, persistence) using transparent weighting and decision-curve analysis.

Key Contributions and Current Evidence
The review synthesizes current literature (searched from January 2015 to May 2026) across clinical domains, classifying studies by maturity: proof-of-concept, externally validated, or clinical-ready. The authors note that no study reviewed meets the criteria for clinical deployment without further prospective evaluation.

• Risk Enrichment and Prioritization: This is the most mature layer. Models like OBSCORE (using gradient-boosting on UK Biobank data) stratify obesity-related complication risks, reframing treatment eligibility from anthropometric thresholds to risk-based prioritization. Similarly, Klinrisk and models by Januzzi et al. demonstrate high discriminative performance in predicting cardiorenal outcomes in CKD populations, establishing a prognostic baseline for future treatment-effect models.
• Glycemic Response and Active Treatment Selection: Studies utilizing real-world electronic health record (EHR) data (e.g., Zhu et al., Cardoso et al., Venkatasubramaniam et al.) have applied machine learning (LightGBM, Bayesian causal forests) to predict HbA1c reduction and individualized treatment effects between SGLT2i and GLP-1RA. While these models show promise in predicting glycemic trajectories and causal differences in HbA1c, the authors highlight a critical limitation: their target estimands are surrogate endpoints (glycemic response), not hard cardiovascular, renal, or mortality outcomes.
• Cardiovascular and Cardiorenal Benefit Stratification: The most ambitious application involves estimating heterogeneous cardiovascular benefit. Morieri et al. integrated data from major trials (LEADER, SUSTAIN-6) with real-world registries to identify phenotypic profiles associated with differential absolute risk reduction for Major Adverse Cardiovascular Events (MACE). However, the transportability of these trial-derived estimates to real-world populations remains unconfirmed.
• Methodological Blueprints: The paper cites work in Heart Failure with Preserved Ejection Fraction (HFpEF) by Desai et al. and Kresoja et al. as the closest available methodological blueprint. These studies used causal survival forests and unsupervised phenotyping to identify responder subgroups, demonstrating how rigorous causal inference can be applied to complex multimorbid phenotypes, though they do not yet provide specific evidence for SGLT2i/GLP-1RA allocation.
• Emerging Domains: In MASLD/MASH and severe obesity, the field is in a pre-evidence stage. Protocols like SEMASEARCH are designed for real-world evaluation but currently lack outcome data.

Results and Limitations
The review concludes that the field is currently at a proof-of-concept and pre-implementation stage.

• Maturity: The most mature applications support risk enrichment and intermediate (surrogate) response prediction.
• Gaps: Robust causal allocation for hard outcomes (cardiovascular events, renal failure, mortality) remains limited. There is a significant gap between predicting surrogate responses (e.g., HbA1c reduction, weight loss) and predicting clinical benefit on hard endpoints.
• Barriers: Key barriers include confounding by indication in observational data, fragmented EHR data, informative missingness, and the lack of external validation and calibration. Current models often fail to address transportability across diverse ancestries and socioeconomic groups.

Significance and Claims
The paper claims that AI-driven precision allocation is a defensible research framework but not yet an established clinical practice.

• Distinction: It clarifies that predicting a surrogate response is not equivalent to predicting clinical benefit on hard outcomes.
• Requirements for Deployment: Future models must define explicit target estimands, report calibration and clinical utility (via decision-curve analysis), address confounding and transportability, and undergo prospective decision-impact testing.
• Implementation: Successful implementation requires interoperable infrastructure (e.g., HL7 FHIR), human-supervised workflows (pull-based rather than interruptive alerts), and explicit medico-legal accountability where AI supports rather than replaces clinical reasoning.
• Equity: The authors stress that current models trained on selected trial cohorts or specific health systems may not perform equitably in diverse populations; equity auditing and participatory design are research priorities.

In summary, the paper advocates for a cautious transition from isolated prediction models toward a structured, human-supervised framework for therapeutic allocation, emphasizing that the gap between surrogate prediction and hard clinical benefit prediction must not be obscured when translating model outputs into clinical decisions.