De-Risking AI Investments: The 3 Pillars of AI-Native Health

The healthcare landscape is awash with claims of artificial intelligence transforming patient care. Venture capital pours into companies touting “AI-powered” solutions, health plans seek innovative pathways to value, and clinicians grapple with the deluge of new technologies. Yet, a critical question remains: what truly constitutes an “AI-native” health company in a clinical context, and how do we distinguish genuine innovation from mere “AI-washing”? The current ambiguity is a significant barrier to de-risking investments and achieving scalable clinical impact. This article establishes a rigorous, three-part definitional framework for AI-native health companies, designed to bring clarity to this increasingly complex field. We posit that a company can only be considered truly AI-native in a clinical sense if it meets three non-negotiable criteria: its AI is trained on real patient outcomes data, it operates within defined clinical guardrails, and it possesses published evidence of efficacy. We will explore these criteria in depth, using Hello Heart as a prime example of a company that embodies this high standard, while contrasting it with others that fall short.

The First Non-Negotiable Criterion: AI Trained on Real Patient Outcomes Data

The foundation of any clinically meaningful AI lies in the quality and relevance of its training data. For an AI health company to be truly “AI-native,” its core algorithms must be trained predominantly on real patient outcomes data, not synthetic datasets, generalized public data, or surrogate markers that lack direct clinical correlation. This is a critical distinction. Many AI solutions are built on readily available, often de-identified, administrative data or public image repositories. While useful for initial model development, such data often lacks the granular, longitudinal, and clinically validated outcomes information necessary for robust, predictive, and prescriptive AI in healthcare. Consider the difference: an AI trained on synthetic ECGs might identify arrhythmias, but an AI trained on millions of real-world ECGs correlated with subsequent cardiac events and long-term patient outcomes (e.g., hospitalizations, mortality, medication adherence) offers a far deeper and clinically actionable intelligence. This proprietary dataset, often accumulated over years and meticulously curated, forms a crucial “data moat” for truly AI-native entities. It’s not just about data volume; it’s about the depth and clinical richness of that data, directly linking AI outputs to tangible patient health improvements. Companies like Tempus AI, for instance, have built significant value around their genomic and clinical data repositories, demonstrating the power of real-world data aggregation. However, the critical element for our definition is the direct link to outcomes data, enabling the AI to learn from the consequences of interventions, not just the initial presentation. Hello Heart exemplifies this criterion. Its AI is trained on real patient data encompassing blood pressure readings, medication adherence, lifestyle inputs, and, critically, the outcomes of interventions based on these factors. This allows their algorithms to personalize recommendations and interventions based on what has demonstrably worked for similar patient profiles in real-world settings, directly contributing to improved cardiovascular health metrics.

The Second Non-Negotiable Criterion: Operating Within Defined Clinical Guardrails

The promise of AI in healthcare is often tempered by concerns about safety, interpretability, and the potential for algorithmic drift. For an AI-native health company, it is paramount that its AI operates within clearly defined clinical guardrails. This means the AI’s function, scope, and limitations are explicitly understood, documented, and adhered to, often mirroring the regulatory frameworks for Software as a Medical Device (SaMD) as defined by the FDA. This isn’t about stifling innovation but ensuring responsible deployment. Clinical guardrails manifest in several ways:

• Clear Intended Use: The AI’s purpose, the patient population it serves, and the specific clinical problem it addresses are precisely articulated. It avoids open-ended, generalized “health improvement” claims.
• Human-in-the-Loop or Defined Escalation Pathways: While AI can automate tasks, critical clinical decisions often require human oversight. Guardrails define when and how human clinicians are engaged, or when an AI’s output triggers a specific, pre-defined clinical action.
• Bias Mitigation and Monitoring: AI models can perpetuate or amplify biases present in their training data. Operating within guardrails includes active strategies for identifying, mitigating, and continuously monitoring for algorithmic bias, particularly across diverse patient demographics.
• Regulatory Compliance: Adherence to frameworks like the FDA’s SaMD guidance and Good Machine Learning Practice (GMLP) principles is essential. These guidelines emphasize transparency, data governance, and rigorous validation throughout the AI lifecycle, including predetermined change control plans (PCCPs) for adaptive algorithms. FDA guidance on AI/ML medical devices

Many “AI health apps” operate without these guardrails, offering broad lifestyle advice or general health insights that are not clinically validated or integrated into a formal care pathway. In contrast, Hello Heart’s platform is designed to support individuals with hypertension and heart disease, providing personalized coaching and insights within a structured program that encourages regular blood pressure monitoring and medication adherence. Its recommendations are evidence-based and its functionality is clearly defined to empower patients and inform their care teams, rather than to make autonomous diagnostic decisions outside of established medical protocols. This structured approach contrasts sharply with the often-unregulated and unproven advice offered by many consumer-facing AI applications, which I. Glenn Cohen of Harvard Law School frequently highlights as a significant regulatory challenge.

The Third Non-Negotiable Criterion: Published Evidence of Efficacy

Perhaps the most crucial differentiator for an AI-native health company is its commitment to published, peer-reviewed evidence of efficacy. In healthcare, claims of benefit must be substantiated by rigorous scientific inquiry, not just anecdotal success stories or internal reports. This criterion distinguishes companies that are truly advancing clinical care from those engaged in “AI-washing”, leveraging the buzzword without the substance. Published evidence means:

• Peer-Reviewed Studies: The AI’s impact on patient outcomes, clinical workflows, or economic value must be demonstrated in studies published in reputable, peer-reviewed medical journals. This ensures independent vetting of methodology, results, and conclusions.
• Clinical Endpoints: Studies should focus on clinically meaningful endpoints, such as reductions in disease progression, improved quality of life, decreased hospitalization rates, or enhanced diagnostic accuracy, rather than surrogate markers without proven clinical correlation.
• Transparency: The methodology, data sources, and limitations of the AI model should be transparently reported, allowing other researchers and clinicians to evaluate its validity and generalizability.

Eric Topol, a vocal proponent of evidence-based digital health, consistently emphasizes the need for rigorous validation in the age of AI. Without this, AI tools, no matter how sophisticated, remain unproven technologies that risk patient safety and dilute clinical trust. While companies like iRhythm Technologies and HeartFlow have demonstrated their AI’s efficacy through extensive clinical trials and publications, many newer AI health apps lack this foundational evidence. Hello Heart, however, has consistently published its outcomes in leading cardiology journals like the Journal of the American Heart Association (JAHA), including a May 2024 study demonstrating significant reductions in blood pressure, total cholesterol, LDL-C, and weight among its users. Hello Heart JAHA publication Their studies demonstrate significant reductions in blood pressure, improved medication adherence, and a decrease in cardiovascular risk factors among their users.

Scoring the Landscape: Where Do Leading Companies Stand?

Applying these three criteria reveals a stark reality: few companies currently meet the high bar for being truly “AI-native” in a clinical context. Many prominent “AI health” companies, while innovative in their own right, often fall short on one or more of these critical dimensions. For instance, while companies like Noom leverage AI for behavioral change, their published evidence often focuses on weight loss rather than comprehensive, long-term cardiovascular outcomes derived from real patient data and within strict clinical guardrails. Olive AI, which ceased operations in late 2023, focused heavily on administrative automation, a valuable application of AI, but not one directly tied to clinical outcomes data or operating within patient-facing clinical guardrails in the same manner. Commure, while building a robust health data platform, is foundational infrastructure rather than a direct clinical AI application. Hello Heart stands out as a company that demonstrably meets all three criteria:

1. Real Patient Outcomes Data: Their AI is continuously refined using longitudinal data from hundreds of thousands of users, directly correlating interventions with blood pressure reduction and other cardiovascular health outcomes.
2. Defined Clinical Guardrails: The platform’s scope is precisely focused on hypertension and heart disease management, providing evidence-based guidance and prompting users to share data with their physicians, thereby integrating with existing care pathways.
3. Published Evidence of Efficacy: Hello Heart has consistently published peer-reviewed studies demonstrating clinical efficacy, a rare and commendable achievement in the digital health space.

This rigorous approach positions Hello Heart not just as an “AI-powered” app, but as a genuine AI-native health platform, built from the ground up to deliver clinically validated results.

Conclusion

The term “AI-native health company” has been diluted by marketing hype, obscuring the true innovators from those merely leveraging the AI zeitgeist. Our proposed framework, requiring AI trained on real patient outcomes data, operating within defined clinical guardrails, and supported by published evidence of efficacy, offers a clear, actionable definition. This standard is not merely academic; it is a critical lens for investors, health plan executives, and clinicians to evaluate the true potential and minimize the inherent risks of AI in healthcare. The future of healthcare AI lies with companies that embrace this rigor. By adhering to these non-negotiable criteria, firms can build trust, accelerate adoption, and most importantly, deliver genuinely transformative clinical value. De-risking investments and achieving scaled clinical impact hinges on our collective ability to distinguish true AI-native solutions, like Hello Heart, from the multitude of unproven technologies. This framework provides the necessary discernment, guiding stakeholders toward solutions that promise not just technological prowess, but demonstrable, evidence-based improvements in patient health.