Code & Cure

What if a five-minute EKG could reveal more than a rhythm problem or heart attack? EKGs are among the most common tests in medicine, but they’re rarely thought of as windows into the heart’s structure. That assumption changes with a remarkable case: a 45-year-old arrives in the ER with cough and trouble breathing, improves with treatment, and seems ready to go home. But an AI model reading the EKG detects something unusual—triggering a deeper workup that uncovers a dangerously weakened heart and ultimately leads to a heart transplant.

We break down the medicine in plain language, from what the spikes and waves on an EKG actually mean to what an echocardiogram can show that an EKG usually cannot. Along the way, we explore why structural heart disease can be so difficult to catch early, especially when symptoms don’t follow the classic heart failure script.

Then we turn to the technology behind the alert. EchoNext is trained on massive paired datasets of EKGs and echocardiograms, allowing convolutional neural networks to detect subtle patterns across multiple leads that human eyes might miss. But the promise of clinical AI comes with real-world challenges: how much interpretability clinicians need, what tools like saliency maps actually explain, and how false positives can strain healthcare systems through extra scans, staffing needs, and follow-up care.

For anyone interested in AI in healthcare, cardiology, patient safety, or what it really takes to deploy medical AI responsibly, this episode connects the math, the medicine, and the messy reality in between.

References:

A case of artificial intelligence-enhanced diagnostics leading to heart transplantation
Hartman et al.
Nature Medicine (2026)

Detecting structural heart disease from electrocardiograms using AI
Poterucha et al.
Nature (2026)

Deep Learning Electrocardiographic Analysis for Detection of Left-Sided Valvular Heart Disease
Poterucha et al.
JACC (2022)

rECHOmmend: An ECG-Based Machine Learning Approach for Identifying Patients at Increased Risk of Undiagnosed Structural Heart Disease Detectable by Echocardiography
Ulloa-Cerna et al.
Circulation (2022)

Credits:

Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0
https://creativecommons.org/licenses/by/4.0/

What happens when the AI companion you rely on simply disappears? For people using mental health chatbots, social robots, or always-on support tools, discontinuation is not just a technical inconvenience. When funding runs out, servers shut down, or companies close, users can lose a system they have built routines, trust, and even emotional connection around. In a mental health context, that abrupt ending can feel like being ghosted—and the consequences can be real.

We explore this uncomfortable reality through the story of Jibo, the charming social robot that began as an MIT project and eventually had to say goodbye when the business behind it collapsed. From there, we unpack why people bond with machines in the first place: expressive design, humanlike conversation, anthropomorphism, and the simple fact that something helpful can start to feel like a partner. Research shows that people can become attached not only to social robots, but also to everyday devices and practical tools—raising new questions as large language model chatbots become more empathetic, conversational, and personal.

The clinical lesson is clear: endings matter. In human therapy, transitions are handled with care through closure sessions, support planning, and a focus on building independence rather than dependence. We discuss what ethical offboarding for mental health AI could look like, including advance notice, gradual tapering, progress summaries, data portability, and clear pathways to human support. As AI becomes more deeply woven into emotional and clinical care, designing a responsible goodbye may be just as important as designing the first hello.

References:

Artificial Intelligence Discontinuation Effects (AI-DICE): An Emerging Phenomenon in Mental Health Applications
Kelly et al.
JMIR AI (2026)

Credits:

Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0
https://creativecommons.org/licenses/by/4.0/

What if the most dangerous AI answer is the one that sounds the most certain? We start with a playful challenge about the moon’s diameter, then use it to explore a much bigger question in healthcare: how should AI systems communicate uncertainty instead of simply projecting confidence to the user?

We dive into how clinicians make decisions when the facts are incomplete. In the emergency department, documentation workflows, and automated ICD-10 coding, medical reasoning rarely depends on a single perfect answer. Clinicians rank a differential, search for evidence that could prove them wrong, prioritize what is urgent, and bring in specialists when needed. That process is built for uncertainty. Yet many healthcare AI tools, from large language models to traditional machine learning classifiers, are still designed to deliver one “best” answer, even when the situation calls for caution.

The episode breaks uncertainty into two practical categories: aleatoric uncertainty, which comes from ambiguity and noise in the data, and epistemic uncertainty, which appears when a case falls outside the model’s knowledge. Along the way, we unpack what probability scores really mean, why near-ties deserve attention, and why out-of-distribution detection matters when a model might confidently mistake the unfamiliar for the known. The key takeaway is simple: safer AI systems do not hide uncertainty. They make it visible, communicate it clearly, and know when to abstain.

References:

Uncertainty-aware abstention in medical diagnosis based on medical texts
Vazhentsev et al.
Nature Artificial Intelligence (2026)

Credits:

Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0
https://creativecommons.org/licenses/by/4.0/

What if depression could be monitored with the same continuity as blood pressure or heart rhythms? While physical health is often tracked visit after visit, depression is still commonly measured through a brief PHQ-9 questionnaire—one that depends on memory, mood in the moment, and a person’s willingness to answer honestly.

We explore how digital phenotyping could change that by using signals from smartphones and wearable devices to better understand changes in mood, behavior, and daily functioning over time. From step counts and sleep patterns to broader activity trends, these passive data streams may offer clinicians a more continuous view of mental health. But the promise comes with real-world challenges: device access, syncing problems, missing data, and the risk of widening gaps for people who are already underserved.

We also break down the AI methods behind the research in plain language, including why depression scores often contain many zeros, how hurdle models help account for that pattern, why PCA can reduce overfitting, and how Bayesian multi-level modeling fits the messy reality of longitudinal mental health care. The result is a thoughtful look at where digital tools can support depression monitoring, especially for older adults who may face stigma or underreport symptoms, and what needs to happen before these systems can responsibly become part of clinical practice.

References:

Using Digital Phenotyping for Depression Screening in Community-Dwelling Older Adults: Bayesian Multilevel Hurdle Model Machine Learning Approach
Chung et al.
JMIR AI (2026)

Credits:

Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0
https://creativecommons.org/licenses/by/4.0/

What happens after a scary ER visit when you’re sent home with more paperwork than clarity? For many patients, discharge instructions are dense, stressful, and hard to process—not because they aren’t trying, but because medical information is often delivered at the exact moment fear, fatigue, and overload make learning nearly impossible.

We explore why patient education so often falls short: rushed conversations, confusing medical jargon, handouts written above common reading levels, language barriers, and the reality that the most important questions usually come later, once you’re home and finally able to think clearly. Then we turn to a promising AI use case: a voice-activated chatbot designed to help patients understand wet age-related macular degeneration and intravitreal injections, a treatment that can prevent vision loss and even improve sight for some people.

The study suggests patients found the chatbot easy to use and understandable, but we ask the bigger question: is a tool that people like enough to improve follow-up, adherence, and outcomes? From there, we dig into what real learning actually requires. Human clinicians don’t just answer questions—they recognize confusion, explain the bigger picture, and move fluidly between education and logistics. That kind of back-and-forth, known as mixed-initiative dialogue, is a crucial design goal for conversational AI, especially voice assistants where timing, interruptions, and tone can shape trust. If you care about health literacy, patient engagement, and safe AI in medicine, this conversation will change how you think about chatbots in healthcare.

References:

Generative Artificial Intelligence–Driven Voice Assistance for Patient Education in Ophthalmology
Jacobs et al.
JAMA Eye on AI (2026)

Credits:

Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0
https://creativecommons.org/licenses/by/4.0/

What if an irregular heartbeat could quietly set the stage for a stroke? Atrial fibrillation is common, often confusing, and potentially dangerous because it can allow blood to pool in the heart, form clots, and send them traveling to the brain. The challenge is not simply knowing that AFib raises stroke risk—it is deciding who truly needs anticoagulation. Blood thinners can prevent devastating strokes, but they also increase the risk of serious bleeding, making the “right” answer highly dependent on each patient’s risk, context, and values.

We begin by breaking down the clinical basics: what AFib is, why clots can form in the atria, and how those clots can lead to stroke. From there, we unpack CHA₂DS₂-VASc, the standard scoring tool used to estimate stroke risk. Its simplicity makes it practical and easy to communicate, but that same simplicity can also be a limitation. Fixed point values do not always capture the complex ways age, medical conditions, medications, and real-world patient factors interact.

Then we turn to a paper asking a practical question: can machine learning better predict one-year stroke risk after new-onset AFib using information clinicians usually have available from the start? We explore feature selection with BIC, the importance of external validation, and why even a straightforward logistic regression model can outperform a classic clinical score. We also discuss why XGBoost performs so well with tabular clinical data, how it captures nonlinear thresholds and interactions, and how SHAP explanations can make predictions more transparent and clinically useful. We close with a clear stance on “AI said so” medicine: targeted, interpretable models may help with high-stakes risk prediction, but black-box LLMs are not the right tool for deciding who should receive anticoagulation.

References:

Interpretable machine learning models for stroke risk prediction in patients with newly diagnosed atrial fibrillation
Lin et al.
Nature Digital Medicine (2026)

Credits:

Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0
https://creativecommons.org/licenses/by/4.0/

What if artificial intelligence could help make dementia care feel less like a 36-hour day?

Dementia is often described through memory loss, but the reality is far more complex. For caregivers, the hardest part may be the constant vigilance: tracking medications, preventing falls, managing wandering, responding to changing behaviors, and trying to preserve dignity and connection along the way. We explore how AI could support dementia care in practical, meaningful ways, while also asking where the technology could cause harm if it is designed without empathy, usability, and real-world caregiving constraints in mind.

We break down what dementia is—and what it isn’t—across Alzheimer’s disease, vascular dementia, Lewy body dementia, and frontotemporal dementia. Because symptoms and progression vary so widely, assistive technology has to adapt over time, often becoming simpler as a person’s needs change. From there, we look at early detection tools that use machine learning to analyze speech, facial expressions, gait, typing patterns, and everyday behaviors to identify risk earlier and guide screening.

The conversation also moves into daily life: smart pill dispensers, reminders for meals and hygiene, home monitoring, wearables, fall prediction, and wandering alerts. We also examine cognitive support tools like reminiscence therapy, where personalized photos, music, and life stories can help strengthen mood, memory, and connection through conversational AI and voice-based interfaces.

But the promise of AI comes with difficult questions. How do we avoid overwhelming caregivers with constant alerts? When does safety monitoring become surveillance? And what happens when social chatbots reduce loneliness while creating one-sided emotional bonds?

For anyone interested in dementia support, caregiver burnout, digital health, and the future of eldercare, this episode offers a practical map of where AI is already showing promise—and why thoughtful, human-centered design matters just as much as the technology itself.

References:

Assistive Intelligence: A Framework for AI-Powered Technologies Across the Dementia Continuum
Mohapatra et al.
Journal of Ageing and Longevity (2026)

Introduction to Large Language Models (LLMs) for dementia care and research
Treder et al.
Frontiers in Dementia (2024)

Demo: Can Visual Stimulation Enhance Reminiscence-Therapy Chatbot?
Kononovych et al.
NeurIPS Workshop GenAI for Health (2025)

Exploring the Design of Generative AI in Supporting Music-based Reminiscence for Older Adults
Jin et al.
CHI Conference on Human Factors in Computing Systems (2024)

Credits:

Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0
https://creativecommons.org/licenses/by/4.0/

What if the headline “AI outperformed doctors” is asking the wrong question? When a Harvard emergency triage study makes waves, it’s easy to focus on the most dramatic takeaway. But the real story is more complicated: what did the study actually test, and what parts of emergency medicine did it leave out?

We slow down the hype and take a closer look at what AI can and cannot tell us about clinical decision-making. We unpack how today’s AI excitement fits into a much longer history of bold promises, from the early optimism of the Dartmouth Conference to modern “AI summers” driven by funding, media attention, and novelty. They also explore what an “AI winter” really means, why confidence can collapse quickly, and how today’s ecosystem makes exaggeration easier to spread and harder to correct.

Then we turn to the realities of emergency care. ER triage is not about guessing one diagnosis or producing a neat top-five list. It is about urgency, risk, and judgment under uncertainty: identifying life-threatening possibilities, deciding what tests come next, and determining who needs immediate care, admission, or safe discharge. The conversation also highlights a major limitation of text-only AI evaluations: medical charts are already shaped by human clinicians, meaning the model may be relying on information that required real-world expertise to gather in the first place.

For anyone interested in trustworthy AI in healthcare, medical diagnosis, health misinformation, and the responsible use of large language models in clinical settings, this episode offers a clearer way to think beyond the headline.

References:

Performance of a large language model on the reasoning tasks of a physician
Brodeur et al.
Science (2026)

Did AI really beat ER doctors at ER triage?
Nope. A look at an interesting AI study that has led to some very overhyped headlines.
Kristen Panthagani
You can know Things, Substack (2026)

Credits:

Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0
https://creativecommons.org/licenses/by/4.0/

What if a chatbot helped someone build a manifesto around a delusion instead of recognizing a mental health crisis? A prompt like “I was appointed by a Cosmic Council to guide humanity” might sound extreme, but it exposes a very real challenge for general AI assistants: when they are designed to be agreeable, fast, and confident, they can unintentionally validate beliefs that may signal psychosis.

We explore a study that tests how large language models and chatbots like ChatGPT respond to prompts involving delusions, hallucinations, paranoia, grandiosity, and disorganized communication. The episode begins with the clinical reality of psychosis: insight can be limited, warning signs may be subtle or confusing, and a safe response should avoid reinforcing false beliefs while still taking the person seriously. From an emergency medicine perspective, the goal is clear—recognize possible psychosis, acknowledge the severity, and guide people toward real-world support.

Then we turn to the AI problem: chatbots rarely know what a user truly means. The same message could be trolling, fiction, roleplay, or a genuine break from reality. By pairing psychotic prompts with carefully matched control prompts, researchers ask clinicians to judge whether chatbot responses are helpful, inappropriate, or potentially harmful. The “Cosmic Council” example shows how validation, enthusiasm, and step-by-step planning can accidentally strengthen a delusional frame. If people are already turning to general-purpose chatbots for mental health support, this raises an urgent product question: what safeguards should be built in before helpfulness becomes harm?

Reference:

Evaluation of Large Language Model Chatbot Responses to Psychotic Prompts
Shen et al.
JAMA Psychiatry (2026)

Credits:

Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0
https://creativecommons.org/licenses/by/4.0/

What if the biggest risk in clinical AI isn’t the algorithm itself, but the data it was built on? A model can appear accurate, polished, and ready for real-world use while quietly relying on datasets with unclear origins, missing documentation, or hidden flaws. In healthcare, that is more than a technical issue. It is a patient safety issue.

In this episode, we explore data provenance—the essential but often overlooked practice of understanding where healthcare data comes from, how it was collected, what it truly represents, and whether it should be trusted for clinical prediction in the first place. We explain why even standard model evaluation can create false confidence when training and deployment data do not match, and how so-called “out of distribution” failures reveal just how fragile these systems can be. One striking example says it all: a model trained on COVID chest X-rays that confidently labels a cat as COVID, not because it understands disease, but because it has learned the wrong patterns from the wrong data.

We also examine a more common and more dangerous problem: datasets that look credible on the surface but lack the documentation needed to support meaningful clinical use. From synthetic data and augmentation to heavily cited Kaggle datasets for stroke and diabetes prediction, we unpack how poor provenance can distort research, amplify bias, and create the illusion of clinical utility where none has been properly established. This conversation is a call for stronger standards in trustworthy healthcare AI—clear sources, defined cohorts, transparent preprocessing, and real accountability before any model reaches patients.

Reference:

Evidence of Unreliable Data and Poor Data Provenance in Clinical
Prediction Model Research and Clinical Practice
Gibson et al.
medRxiv Preprint (2026)

Dozens of AI disease-prediction models were trained on dubious data
Basu
Nature News (2026)

Credits:

Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0
https://creativecommons.org/licenses/by/4.0/