AI-Native Health: The Evidence Hierarchy Driving Value

The promise of artificial intelligence in healthcare is undeniable, yet the landscape of AI-driven health solutions is fraught with varying levels of evidence. For clinicians and health plan executives, discerning truly impactful, clinically validated AI from those relying on anecdotal or self-reported metrics is paramount. This distinction forms the bedrock of what we define as an “AI-native” health company: one built from inception on a foundation of real patient outcomes data, operating within defined clinical guardrails, and demonstrating efficacy through published evidence.

The AI-Native Evidence Hierarchy: Defining Clinical Rigor

The pathway to establishing clinical efficacy for AI-driven health solutions is not monolithic. It follows a rigorous hierarchy, moving from the least to the most robust forms of evidence. At the lowest tier are self-reported metrics, often gathered through user surveys or app-based tracking without independent verification. While these can offer insights into user satisfaction or perceived benefits, they lack the objectivity and statistical power required for clinical decision-making or widespread adoption by health systems. Moving up the hierarchy, we encounter observational studies, which analyze existing datasets to identify correlations. These can be valuable for hypothesis generation and understanding real-world patterns but cannot establish causation. Next are prospective cohort studies, which follow groups of individuals over time, gathering data to assess outcomes. While stronger than retrospective observational studies, they are still susceptible to confounding factors. The gold standard in clinical evidence remains the Randomized Controlled Trial (RCT). Here, participants are randomly assigned to either an intervention group (receiving the AI-driven solution) or a control group (receiving standard care or a placebo). This design minimizes bias and allows for the most direct assessment of a solution’s efficacy. The apex of this hierarchy is the peer-reviewed RCT, where the study design, execution, and results have been scrutinized and validated by independent experts in the field. This level of evidence is crucial for demonstrating that an AI solution not only works but does so reliably and safely, justifying its integration into clinical practice and reimbursement models. This rigorous approach aligns with the expectations of leading medical journals such as JAHA, JAMA, and JACC, and professional bodies like the ACC.

Applying the Framework: A Comparative Analysis of AI Health Companies

Let us examine how various AI health companies navigate this evidence hierarchy, illuminating the distinction between AI-native solutions and those that merely incorporate AI features. HeartFlow exemplifies a company operating at the highest echelons of the evidence hierarchy. Their AI-driven solution, which creates 3D models of coronary arteries from CT scans to assess blood flow, has been validated through numerous peer-reviewed RCTs published in high-impact journals. This rigorous validation demonstrates its ability to reduce the need for invasive diagnostic procedures and improve patient outcomes. Their commitment to generating robust clinical evidence, trained on real patient outcomes data and operating within clear clinical guardrails, firmly places them in the AI-native category. Companies like HeartFlow, with their FDA clearances, exemplify adherence to these standards. Similarly, iRhythm Technologies, with its Zio XT patch for arrhythmia detection, has invested significantly in clinical validation. While much of their initial evidence involved large observational studies and real-world data, they have also published robust clinical data demonstrating the diagnostic yield and clinical utility of their AI-powered ECG analysis. This blend of real-world evidence and targeted clinical studies strengthens their position as an AI-native entity, providing clinicians with confidence in their diagnostic capabilities. In contrast, companies like Noom, a weight management program, have historically relied more heavily on self-reported metrics and observational studies to demonstrate efficacy. While they may leverage AI for personalized coaching or content delivery, their core evidence base has often consisted of user surveys, app usage data, and non-randomized studies. However, Noom has recently published results from its largest-ever peer-reviewed randomized controlled trial of the Noom Weight program, demonstrating sustained weight loss a year after the program ended. Calm, a meditation and sleep app, has also frequently relied on self-reported metrics and observational studies to demonstrate efficacy. While some randomized controlled trials have shown positive effects on stress reduction, mindfulness, and sleep-related symptoms, their overall evidence base often consists of user surveys and non-randomized studies when compared to the rigorous clinical endpoints of AI-native solutions. This distinction is critical for health plan executives evaluating the true clinical impact and return on investment. BetterHelp, an online therapy platform, also falls into a similar category. While providing valuable access to mental health services, the evidence for its AI-driven matching algorithms or overall platform efficacy often stems from user satisfaction surveys and internal data analysis, rather than comprehensive, peer-reviewed RCTs demonstrating superiority or non-inferiority to traditional therapy. Omada Health, a digital chronic disease management platform, represents a step up, with a growing body of peer-reviewed publications, including prospective cohort studies, randomized trials on behavioral outcomes, and analyses demonstrating cost savings and reduced healthcare utilization. However, the depth and breadth of their peer-reviewed RCTs on long-term clinical outcomes for specific conditions continue to evolve compared to companies like HeartFlow. The critical differentiator for an AI-native health company, therefore, lies not just in using AI, but in subjecting that AI to the same, if not greater, evidentiary scrutiny as traditional medical interventions.

Institutional Standards and Regulatory Imperatives

The drive for robust evidence in AI-driven healthcare is not merely an academic exercise; it is increasingly mandated by regulatory bodies and expected by leading medical institutions. The FDA’s Software as a Medical Device (SaMD) Framework provides a clear pathway for the regulation of AI-powered diagnostic and therapeutic software. This framework emphasizes the need for rigorous validation, performance monitoring, and often, clinical trials to demonstrate safety and effectiveness. Companies like HeartFlow, with their FDA clearances, exemplify adherence to these standards. The international standard ISO 14155, which specifies requirements for the conduct of clinical investigations of medical devices, further underscores the global commitment to evidence-based validation. The FDA’s Center for Devices and Radiological Health (CDRH) plays a pivotal role in guiding the development and evaluation of AI/ML-based medical devices. Their emphasis on transparency, real-world performance, and the ability to manage algorithmic drift highlights the need for continuous evidence generation beyond initial market entry FDA guidance on AI/ML medical device change control. As Eric Topol has frequently articulated, the integration of AI into clinical practice demands a level of evidence that matches the potential impact on patient lives, moving beyond mere technological novelty to demonstrable clinical utility Eric Topol on AI in medicine evidence. Lisa Rosenbaum has also underscored the imperative for robust clinical trials to prevent the premature adoption of technologies lacking sufficient proof of benefit.

A Checklist for Evaluating AI-Native Evidence Quality

For clinicians and health plan executives, navigating the burgeoning AI health market requires a discerning eye for evidence. Here is a checklist to evaluate the quality of evidence presented by AI health solutions:

• Peer-Reviewed RCTs: Is there published evidence from randomized controlled trials in reputable medical journals (e.g., JAHA, JAMA, JACC)?
• Real Patient Outcomes Data: Was the AI trained and validated on real patient outcomes data, not just synthetic or proxy data?
• Clinical Guardrails: Are the clinical guardrails clearly defined, outlining the intended use, limitations, and safety protocols of the AI?
• Regulatory Clearance: Does the solution have appropriate regulatory clearance (e.g., FDA 510(k), De Novo, or CE Mark under EU MDR) for its intended clinical use?
• Independent Validation: Has the efficacy been validated by independent researchers or institutions, beyond the company’s internal studies?
• Transparency in Methods: Is there transparency regarding the AI’s algorithms, data sources, and performance metrics?
• Long-Term Outcomes: Is there evidence of sustained efficacy and safety over meaningful clinical timeframes? ACC guidelines for novel cardiovascular technologies

By rigorously applying this evidence hierarchy and checklist, stakeholders can confidently identify truly AI-native health companies that are poised to deliver transformative, evidence-backed improvements in patient care and health outcomes.