Predict, Prevent, Personalize: Medicine Reimagined Through AI

Predict, Prevent, Personalize: Medicine Reimagined Through AI

Artificial intelligence in medicine is not a recent phenomenon. It has been a part of clinical research and practice since the earliest days of the field itself. What has changed is the scale, the sophistication, and the urgency with which health systems are now deploying AI tools across clinical and operational settings. Dr. Theodore Zanos, Head of the Division of Health AI at Northwell Health and the Feinstein Institutes for Medical Research, presents the full breadth of how AI is being applied to transform patient care in a session on 1ArtificialIntelligence hosted by Gil Bashe, Chair of Global Health and Purpose at FINN Partners, spanning everything from predicting which hospital patients will deteriorate hours before clinical signs appear to building the most detailed digital twin of the human vagus nerve in existence.

The session covers the history and current taxonomy of medical AI, Northwell's approach to in-hospital deterioration prediction using clinical wearables, the persistent challenge of model drift, an AI-driven solution to the nurse staffing crisis, and the frontier of bioelectronic medicine, where AI-guided nerve stimulation is opening the door to truly personalized therapies.

AI in Medicine Long Before ChatGPT

Zanos opens with a deliberate reframing of the AI conversation. He notes that in the current environment, AI has become essentially equated with large language models and chatbots, and he wants to correct that perception. AI has been around long before ChatGPT. He traces the lineage back to the 1960s, citing Norbert Wiener, the father of cybernetics, as one of the earliest voices to raise concerns about the unchecked application of this type of technology. Medical AI, in particular, has been present since the beginning of the field. One of the earliest AI applications was ELIZA, a chatbot developed at MIT by Joseph Weizenbaum that simulated conversation with a human therapist. That was followed in the 1970s by diagnostic models, and since the 1980s, neural networks and machine learning approaches have been applied across virtually every field of medicine.

The scale of this history is reflected in a striking regulatory data point. The FDA has approved more than 1,200 AI algorithms for healthcare applications to date, and the number grows daily. However, Zanos emphasizes a critical distinction: as of March 2026, zero of those approved algorithms rely on large language models. The FDA is still working to determine how to regulate, validate, and ensure the reliability of LLM-based tools as they enter the healthcare space aggressively. This is not something that we can just turn on and use and trust, Zanos says. It really needs to be regulated and validated properly.

The Spectrum of Medical AI: From Predictive Models to Agentic Systems

Zanos outlines a taxonomy of AI tools currently operating in healthcare, organized by increasing complexity and risk. At the top sits predictive AI, which is the primary focus of his division's work: machine learning models that use historical data, electronic health records, and other modalities to estimate clinical risk or diagnose specific conditions. Below that sits generative AI, which produces clinical text, explanations, and draft discharge summaries. Then come clinical AI assistants, tools that provide services to doctors including chart summarization, inbox triaging, results triaging, drafting responses to patients, and answering clinical queries, as well as the ambient AI technologies that capture and document clinical encounters.

At the frontier is agentic AI, which automates multiple steps across multiple agents performing specific tasks. Zanos notes that agentic AI has not yet seen prime time in healthcare, pointing to a single health system in Utah that has implemented an agentic approach for prescriptions as the leading edge. The important observation is that each category carries different risk profiles, different requirements for clinician involvement, and different governance needs. Moving from predictive models toward agentic systems increases both the potential capability and the potential for harm.

Preventing In-Hospital Deterioration: A Three-Step Approach

The central clinical problem Zanos addresses is in-hospital patient deterioration, a common but often overlooked challenge. Between 5 and 15 percent of patients admitted to regular medical surgical floors experience deterioration events, which can include rapid response team calls, unplanned ICU transfers, cardiac arrest, intubation on the floor, or death. The root cause, in many cases, is under-triaging from the emergency department, where patients who should have been admitted to an ICU are instead sent to a lower-acuity setting.

Zanos describes a three-step approach to this problem. The first step addresses the opposite end of the acuity spectrum: identifying perfectly stable patients who do not need to be woken at 3 or 4 a.m. for routine vital sign checks. The AI tool his team has developed can identify these patients and potentially reduce overnight wake-ups by 50 percent, improving patient sleep, accelerating discharge, improving outcomes, and enhancing patient experience. The key design principle is that it asks nurses to do less, not more.

The second step focuses on identifying the patients who are most at risk. Northwell's deterioration prediction model uses first labs, demographics, and vitals measured in the emergency department to predict combined outcomes including in-hospital death, unplanned ICU transfers, and rapid response team calls. The model is trained on a large number of hospitalizations, reflecting the scale advantage of operating within the largest health system in the Northeastern United States. It outperforms the widely used Epic Deterioration Index by approximately 15 to 20 percent, and performs consistently across different demographics and across all of Northwell's hospitals. The model is currently running silently across several hospitals, with clinical deployment planned for the summer or fall.

The third step deploys clinical wearables for the highest-acuity patients identified by the model, replacing manual vital sign checks with continuous monitoring at clinical-grade accuracy.

Clinical Wearables and the 17-Hour Warning Window

Northwell has tested three different clinical wearable devices and deployed patches to thousands of patients across its hospitals. An ongoing multi-year NIH grant is expanding the dataset to approximately 10,000 patients and more than 25,000 patient days, which Zanos says will be the largest clinical wearable in-hospital dataset in the country.

The advantage of continuous wearable data is granularity. These devices monitor vital signs every 5 to 10 seconds, compared to the intermittent manual measurements that are the current standard. That granularity allows AI models to detect subtle changes in vital signs, telltale signs of deterioration that would be invisible in intermittent data. The published research from Zanos's team demonstrates that algorithms trained on wearable data can predict deterioration events on average approximately 17 hours in advance, a lead time that is clinically significant. If you tell a clinical team that somebody is going to deteriorate in the next half hour, there is very little they can do, Zanos explains. But if you tell them early enough that signs are not directly visible to them, this could initiate actions that could potentially alleviate the deterioration event.

An important technical finding is that the algorithm is device agnostic. The team trained the model using data from one device and tested it using data from a different device. As long as the devices are accurate, the model works equally well, which is a meaningful feature for scalability across health systems that may use different wearable platforms.

Clinical Phenotyping: Understanding Not Just Who, But Why

Zanos argues that prediction alone is insufficient. If an AI system tells a provider that a patient is going to deteriorate without indicating why, the provider is left without actionable guidance beyond taking another look at the patient. His team is working to embed phenotyping into the prediction pipeline, using unsupervised learning approaches to identify distinct clinical phenotypes of deterioration and to assign individual patients to those phenotypes in real time.

The precedent for this work comes from COVID-19 research conducted at Northwell, where the team identified four distinct clinical phenotypes among hospitalized patients, including renal and high-inflammation lung profiles. These phenotypes had meaningfully different mortality rates and disease trajectories, and they shifted in prevalence across different variants and waves. Critically, individual patients could switch phenotype membership during a single hospitalization, meaning that the underlying cause of potential deterioration could change over time. The goal is to provide clinicians not just with the information that a patient is at risk, but with insight into the specific physiological trajectory driving that risk, enabling more targeted and timely interventions.

The Challenge of Model Drift in Healthcare AI

One of the most technically consequential sections of the session addresses model drift, the tendency of AI models to lose performance over time or when deployed in a different setting from where they were trained. Zanos presents this as a structural reality of healthcare AI, not an exception. Patients change, new treatments emerge, viral variants shift, and clinical practices evolve. A model developed in a specific place at a specific time will likely have decreased performance later or elsewhere.

To quantify this, Zanos references a forthcoming study in which his team gathered all published external validations of five clinical decision support tools built into Epic, the largest electronic health record platform in the United States. Across 22 studies, 32 sites, and more than 3 million patients, the models showed consistent performance degradation. The sepsis model, for example, demonstrated an AUROC of 0.65 in real-world validation, down from the 0.8 reported by Epic. The pattern held across readmission and other models as well.

Northwell experienced the same phenomenon with its own COVID mortality models, where performance drifted rapidly as conditions changed. The response was to develop a self-monitoring and auto-updating framework that allows models to detect their own performance drops and retrain automatically when needed. Zanos frames this not as a solved problem but as a foundational requirement of responsible AI deployment in healthcare: models must be continuously monitored, and health systems must develop the infrastructure to detect and correct drift.

Operational AI: Solving the Nurse Staffing Crisis

Zanos shifts from clinical to operational AI to address a problem with equally significant implications for patient care: the nurse staffing crisis. Career satisfaction among nurses has been dropping steadily, and the exodus from the profession has created a vicious cycle. As nurses leave, the remaining staff absorb additional workload, leading to increased burnout and further attrition. Every departure triggers a hiring process that can take months from application review through onboarding and training, during which the gap is filled either by increasing overtime for existing nurses or by hiring expensive temporary flex staff.

Northwell's approach is preemptive rather than reactive. The team has adapted Amazon's DeepAR forecasting framework to build a combined demand and attrition model. The demand model predicts how many nurses will be needed 6 to 12 months in the future. The attrition model predicts how many nurses are likely to leave during the same period, using data including historical demand and attrition patterns, demographic data, role and job-related data, and commute distance. The combination produces a hiring recommendation that tells nurse managers to begin recruiting now, before vacancies materialize, so that new hires are ready when they are needed.

The financial impact is substantial. Estimated cost savings from deploying the model across just 10 of more than 400 eligible units amount to approximately 10 million dollars per year, driven primarily by reduced reliance on expensive flex staffing. Beyond the cost dimension, the team is also applying the same phenotyping logic used in clinical deterioration to understand why nurses leave, identifying modifiable factors that could inform better retention strategies.

Bioelectronic Medicine and the Vagus Nerve Digital Twin

The final research vertical Zanos presents is the most forward-looking: the intersection of AI and bioelectronic medicine, a field pioneered at the Feinstein Institutes and Northwell Health. Bioelectronic medicine uses electrical stimulation of specific nerves, most prominently the vagus nerve, to treat conditions ranging from epilepsy to rheumatoid arthritis. There are approximately 300 active clinical trials in the space, and the FDA recently approved a vagus nerve stimulation device for rheumatoid arthritis, a milestone for the field.

The vagus nerve is the tenth cranial nerve, connecting the brain to the heart, lungs, intestines, liver, spleen, and most other peripheral organs. Each side contains approximately 100,000 nerve fibers arranged in highly complex fascicle structures that twist, merge, and split along the length of the nerve. Current bioelectronic medicine devices wrap an electrode around the entire nerve and stimulate all fibers simultaneously. This one-size-fits-all approach limits efficacy and produces side effects, because the stimulation affects functions the therapy is not targeting.

To enable truly personalized stimulation, Northwell is building the most detailed digital twin of the human vagus nerve in existence. The project involves more than 3 million micro-CT images and over 4,000 histological images, comprising approximately 200 terabytes of data. Custom deep learning algorithms separate individual fascicles and fibers, creating a three-dimensional map that Zanos compares to the New York subway system: a metro map of the different lines, the different stops, the different organ terminals, so that clinicians know which line to stimulate and where to get off to achieve the most optimal personalized electronic therapy.

Related work includes building algorithms to predict the efficacy of vagus nerve stimulation for epilepsy by integrating MRI, EKG, and EEG data, and developing biomarkers for PTSD using physiological signals, brain waves, and heart rate variability to identify the presence of the condition in populations where underdiagnosis is a significant problem.

Responsible Deployment: Real Problems, Real ROI

Zanos closes with a set of principles that frame the division's philosophy. The motto of the Division of Health AI is Deploy Researched AI and Research Deployed AI. The first half means that tools validated through peer-reviewed research should be pushed toward real clinical deployment rather than remaining as published papers. The second half means that when external AI tools are adopted by the health system, they must be rigorously validated in the local environment to ensure they work for Northwell's clinicians and patients.

He makes a pointed observation about the organizational realities of AI deployment. Clinicians are the users. If they decide not to use a tool, no matter how important the developers believe it is, no AI FOMO will save it. Clinical alerts, for example, can add stress to already overburdened providers if not designed with their workflows and cognitive load in mind. Companies will promise a lot of things, Zanos notes, but even accurate models may fail to deliver value if the integration into clinical workflows is poorly designed or if providers do not trust the system.

Bashe asks Zanos what other health systems, particularly those without the scale or resources of Northwell, should do to build AI capabilities. Zanos responds with an important clarification: he is not a clinician. His background is engineering, and he holds a PhD in biomedical engineering. He works closely with clinicians on his team and across Northwell, and emphasizes that working with clinicians is the only way to be successful in this space, because they provide the necessary clinical insights to identify the right questions to ask and to navigate the pitfalls of development. But the distinction matters for the broader question, because it underscores that health systems do not need to develop AI tools internally. External solutions exist and are appropriate in many cases. Northwell itself is not a tech company or an AI company but a healthcare system, and it uses many external AI tools alongside what it builds in-house. What every institution does need, however, is the capacity for oversight and validation: the ability to assess whether a technology works for their clinicians and their patients, whether it fits into workflows, whether it performs as advertised, and whether it continues to perform over time. You don't need to develop the expertise to develop these technologies, Zanos says, but you do need to develop the oversight and the validation capabilities.