Jan 2, 2019 · 29m · a16z
a16z Podcast | Putting AI in Medicine, in Practice
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the a16z podcast, host Hannah and industry experts explore the practical integration of artificial intelligence into clinical medicine, discussing technical capabilities, reimbursement incentives, and operational strategies for successful adoption.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 2.1 Guest teaching 4.0 Guest disagreement 1.9 The host pushing back 1.4
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 15:00 Vijay rejects Mintu's definition of machine learning
Vijay directly interrupts and corrects Mintu's assertion that machine learning is by definition statistical overfitting, clarifying that proper ML actively avoids overfitting through regularization.
Hardest push from the host ▶ 9:36 Host challenges Mintu on AI versus data collectionHannah pushes back when Mintu dissents on AI's necessity for early detection, demanding to know how his view differs from theirs and why earlier detection wouldn't be inherently beneficial.
Biggest teaching moment ▶ 2:07 Brandon explains why superior AI systems fail to deployBrandon educates the host on system-level deployment barriers, using the 1978 MYCIN case study to show how technical superiority over physicians fails to translate to hospital adoption without financial reimbursement alignment.
The host holds their own ▶ 3:18 Host presses financial logic behind misdiagnosis incentivesHannah challenges Brandon's claim about hospital incentives by highlighting the moral and professional contradiction that no doctor or hospital actually wants incorrect diagnoses, forcing him to clarify fee-for-value nuance.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| History of Medical AI and System Incentives | 2 | 4 | 1 | 2 | The host facilitates a discussion on historical AI medical systems like MYCIN and financial incentives in healthcare. She asks helpful follow-ups and mildly questions how misdiagnosis incentives work in practice, while the guests explain system reimbursement structures. | |
| Diagnostic Applications and Continuous Wearable Data | 2 | 3 | 1 | 1 | Guests outline diagnostic applications, continuous wearable monitoring, and primary care access gaps. The host tracks the conversation, synthesizing key insights such as missing data acting as a health indicator. | |
| AI Autonomy Human Error and Liability | 3 | 5 | 3 | 3 | Mintu offers a counter-perspective that data collection rather than AI is the primary hurdle for early detection. The host presses him on how his view differs, leading into a discussion on AI recapitulating human error and levels of autonomy. | |
| Multi-Modal Data Synthesis and Model Overfitting | 2 | 5 | 5 | 1 | A notable technical debate occurs between guests when Mintu defines machine learning as statistical overfitting and Vijay directly corrects him. The host helps unpack the concept of generalization using the guest's classroom analogy. | |
| Digital Health Research and System Adoption | 2 | 3 | 1 | 1 | Brandon details the high dropout rates in mobile health studies and the necessity of interdisciplinary teams. The conversation is collaborative and informational as the host prompts questions on system incentives. | |
| AI Model Versioning and Deployment | 2 | 4 | 1 | 1 | Mintu raises a technical question regarding continuous learning versus batch versioning in regulated health AI. Vijay and Brandon explain holdout validation sets and speech recognition deployment practices. | |
| Quality Improvement and Clinical Decision Support | 2 | 4 | 1 | 1 | Mintu explains how AI can assist quality improvement in EKG reading and streamline primitive operating room scheduling workflows. The host reacts to hospital scheduling practices and concludes the interview. |
Statements from this episode (20)
Ballinger: Stanford's 1978 MYCIN AI system outperformed five pathologists
“An interesting case study is, is the Meissen system, which is from 1978, I believe, and so this was an expert system trained at Stanford. It would take inputs that were just typed in manually, and then it would essentially try to predict what a pathologist wou…”
Ballinger: Misdiagnosis generates more hospital revenue under fee-for-service models
“If you think about kind of a hospital from the CFO's perspective misdiagnosis actually earns them more money. Because when you misdiagnose, you do follow-up tests, right? And those, and our billing system is fee-for-service. So every little test that's done is…”
Ballinger: Fee-for-value payment models will enable medical AI adoption
“Things like fee for value are interesting because now you're paying people for, say, an accurate diagnosis or for a reduction in hospitalizations, depending on the exact system, and so I think that's the case where actually accuracy is rewarded with greater pa…”
Ballinger: Doctors cannot diagnose sleep apnea from Fitbit data alone
“No doctor right now can read your Fitbit data and tell you whether you have a condition like sleep apnea.”
Turakhia: AI cannot predict acute heart attacks days in advance
“You can predict a cumulative probability, like a probability of getting condition X or diagnosis X over a time horizon of five or 10 years. But we are nowhere near saying, you know, you're going to have a heart attack in the next three days.”
Turakhia: Missing wearable data is the strongest predictor of illness
“In fact, the biggest predictor, Of someone getting ill with a lot of wearable studies is missing data because they were too sick to wear the sensor.”
Ballinger: Most Americans do not get an EKG until age 65
“Most people actually in the U.S., you get your first D.K.G. When you turn 65 as part of your Medicare checkup.”
Ballinger: Only half of Americans have a primary care physician
“And if you look at that in aggregate, about half of people in the U.S. Have a primary care physician at all, which seems astonishingly low, but that's kind of the fact.”
Ballinger: Large percentages of common medical conditions remain undiagnosed
“About a third of people with diabetes don't realize they have it. About a fifth of people with hypertension. For AFib, it's 30 or 40%. For sleep apnea, it's like 80%.”
Turakhia: Neural networks replicate human error patterns in EKG and imaging studies
“Some of the most promising aspects of the imaging studies and the EKG studies are that the confusion matrices, the way humans misclassify things is recapitulated by the convolutional neural networks.”
Turakhia: Fully autonomous medical AI faces societal, not technical, barriers
“That's a societal issue. That's not a technical hurdle at this point.”
Ballinger: Wearables will generate two trillion data points in 2017
“Wearables are an interesting case because they'll generate about Two trillion data points this year.”
Turakhia: Clinical documentation terminology varies between individual hospitals
“In natural language processing that's embedded in AI, the lexicon that people use, how doctors and clinicians write what it is that they're seeing with their patient is different from not even specialty to specialty, but hospital to hospital, sort of mini subc…”
Ballinger: Wearable device data like Fitbit data is globally standardized
“I think that's actually a nice thing about wearable data is that Fitbits are the same all over the world.”
Ballinger: Early Apple ResearchKit apps lost 90% of participants in 90 days
“The challenge that the first five ResearchKit apps had is that they got 40,000 people, and then they lost 90% of them in the first 90 days.”
Turakhia: FDA's Digital Health Office effectively mitigates regulatory risk
“The regulatory risk thing is being largely addressed by this new Office of Digital Health and the FDA, and they're really doing, seem much more forward thinking about it.”
Turakhia: Continuous learning in medical AI risks patient harm from biased data
“Bad data could heavily bias the system and cause harm, right? So if you start learning from bad inputs that come into the system for whatever reason, you could intentionally or unintentionally, you know, cause harm.”
Turakhia: No standardized quality improvement metrics exist for EKG interpretation
“There's actually no standardized metrics for QI in any of this.”
Ballinger: Full-stack healthcare provider startups will achieve fastest adoption
“I think that's a model that we're going to see probably get the quickest adoption.”
Ballinger: Healthcare will reorganize vertically around specialized AI data networks
“So this is a case where probably we'll see the healthcare industry maybe reconstitute itself by vertical. With AI-based diagnostics or therapeutics, because if you think right now providers are geographically structured, but with AI, every data point makes the…”