Alignment in Medical AI. Clinical Safety and Risk of Institutional Bias

Introduction

As my team continues to operate on the frontiers of Artificial Intelligence (AI) and healthcare, the question of AI alignment continues to remain an alarming one. The crux of AI’s integration into healthcare hinges on alignment — a term that emphasizes the necessity for AI systems to act in accordance with human values and ethics, especially in sensitive sectors like healthcare. The concept of AI alignment in healthcare needs to be revisited repeatedly by those of us operating in the field to answer the pivotal question: How do we achieve medical alignment, and whom do we entrust with this critical responsibility?

Medically Aligned AI

Discussions with my physician peers, who are developing AI in healthcare, we are settling on the concept of medically aligned AI — a specialized approach to AI models that emphasizes personalization and physician attribution.

Unlike current AI development in healthcare applications, medically aligned AI advocates for personalized AI training, such that a unique model is developed for each physician. This goes beyond the current training cycles which are merely overseen with guidance of anonymous doctors. This method of alignment ensures that every piece of advice or diagnosis the model provides is directly attributed to the individual physicians. This results in not one monolithic medical model for all, but many models for many doctors.

This idea starkly contrasts with other models of alignment where the focus might inadvertently shift from patient care to other institutional stakeholders, including corporate, government or medical institutions, whose primary objectives may not align with the personalized needs of the patient.

Framework for Developing Medically Aligned AI

Medically aligned AI is defined by two core principles: the publicly published involvement of licensed medical professionals in model training and the direct attribution of its outputs to individual physicians.

While medical institutions play a crucial role in the dissemination of knowledge and the training of healthcare professionals, their contribution lacks the personal touch and direct accountability inherent in physician-patient relationships. On the other end of the spectrum, corporate-directed AI training, influenced by the imperatives of profit and market share, often misaligns with the fundamental goal of patient care, treating individuals as consumers rather than patients. Likewise, government coerced medical practice has become all too real to those of us practicing women’s health, and AI physician researchers are cautious and wary of this influence embedded in generic publicly available AI models.

Personalization Creates Transparency and Trust

The medically aligned AI principles advocates for physician personalization rather than institutional personalization. Personalized models allow a physician’s to be paired with their AI model. This personalization also allows personality mimicking by the AI. The seamless extension enables a believable and reliable model extending trust of the patient-physician relationship. Personalization also creates a deeply human-centric AI model which accepts that variations of opinion of each physician not only exist, but are healthy and desirable. It acknowledges and decidedly leans into the idea that the human physician’s are inherently variable and can be opinionated. Their interpretation of evidence based literature combined with experience driven decisions are not only appropriate, it is desired and improves the human connection patients seek.

Patients are more likely to trust and follow medical advice when they know it originates from their doctor, rather than an impersonal algorithm or institution. This direct attribution not only enhances the credibility of AI-powered recommendations but also strengthens the bond between patients and their physicians.

In Practice: Bringing Theory to Life

Real-world applications of medically aligned AI, though in nascent stages, offer a glimpse into successful personalization that can be achieved at scale. Several teams, including ours, have developed the technology for rapid cycle customized AI models for physicians. Our soon to be published data from our team and others is imminent on the ability to create usable, accurate, and attributable and personalized AI models.

From AI systems that accurately predict patient deterioration hours before conventional methods to algorithms that tailor rehabilitation exercises to the specific recovery trajectories of patients, the potential is immense. These examples not only validate the feasibility of medically aligned AI but also highlight its profound impact on patient care and outcomes.

Conclusion

Medically Aligned AI

Aligning AI with human values in medicine must prioritize personalization and physician attribution, ensuring outputs are attributed to specific physicians, enhancing personal accountability and patient trust.

Personalization and Trust

Leverage physician-specific AI models to mirror personal and professional characteristics, fostering a stronger patient-physician relationship through trusted, personalized AI advice.

Achievable?

Yes! Technology is now available which can successfully train personalized AI models for physicians.