Medical Doctor, Machine Learning Engineer
Joined August 2017
- Tweets9.2K
- Following724
- Followers1.7K
- Likes80.3K
I was with my senior Registrar when a nurse(an intern) called to Inform me of an emergency. I put the phone on speaker.
I pressed for further details: what's the temperature and Pulse? How's the Child now? She replied "I don't know"
My SR Saw the frustration on my face and said "That how's you people (House Officers) sound every time you call us without details.
I burst out laughing, because I've been guilty of this and didn't know how annoying it can be.
Anyways baby was fine, she just wanted me to come and cannulate.
I understand it's your choice, but here's the fact: female OBGYN residents are fewer than the males.
Where I trained, the ratio was about 1:3, and where I'm doing my housejob it's even worse. It's simply a specialty most female doctors just don't want to go into.
Of the ten or so female friends in my housejob group, I only know one who is writing primaries for Obstetrics and Gynaecology next year. Most guys I know here, in Housejob want to do surgery or Obstetrics. I would, myself, likely write primaries there next year.
In the context of a society like Nigeria, where most women would rather do specialties that afford them more free time (the most common I get from my friends is "omoh, I no get energy for that stress of OBSGYN), you're far more likely to meet a male obstetrician in the hospital on cal
Which means that if you came in with, say, a placental abruption and no female OBGYN was on call that night, nobody would take you seriously if you insisted on one. They'd just leave you alone.
I understand the hesitation you have about male doctors not understanding you, but X no be real life. There are compassionate, competent male Obstetricians who would attend to your complaint and make your childbirth journey pleasant,
But if you insist on working with only females, it would be a really really difficult journey.
"Own your intelligence. Anyone that opposes this has something to sell you"
I will always support open source ❤️
Own your intelligence.
Anyone that opposes this has something to sell you (perhaps your own intelligence back to you.)
Local models need to continue proliferating. The “absolute unit” massive models may be out of reach for now locally, so your best option is to, indeed, keep your options open. Own your harness, data, connections, runtime: your agents. Embrace open source.
Again, I want to be a really really good Medical Doctor. I also want to be a really really good Engineer.
I'm so convinced that there's an intersection between both, and I'll work very hard on both fields.
When opportunity comes, I won't be unprepared
I had to pause this cause because of my Pediatrics postings.
Figured I'd start from the beginning and speed run it up to where I stopped.
👏🏾👏🏾👏🏾 Incredible work brother!
Proud of you and your team.
We are rolling out local language features in the new, upgraded DermAfricans website.
Please help us confirm that our phrasing is correct by reviewing at review.dermafricans.com.
Thank you!
Chad 4 changes everything!
I've had early access, and it blew my mind
A comparison with Opus 5.5 and GPT 6.1 🧵
Introducing Chad 4, the newest frontier model.
It’s 69T Params, scored 120 on AA and surpasses all benchmarks everywhere.
We’ve rolled it out to our trusted mythical partners and a wider rollout shall follow suit.
Input: $0.069/m
Cache: Free
Output: we pay you $5/m
The world’s first cash-positive model. The open weights will also be available soon.
It’s all coming soon! Very soon!
Running some random open source model on my barely alive, gasping , begging me to stop, overheating laptop
I think I'll go ahead and build this. I'll pick a Local Model, fine tune it on the principles behind then MAI Diagnostic Orchestrator.
Then I'd also build an EHR that functions effectively running this.
Yes, I'd open Source the whole project, should be fun,
A 52-year-old man is admitted to a hospital with fever, nausea and breathlessness that has worsened until he needs oxygen.
What investigations do you order next?
Researchers at Microsoft did a study comparing physicians with artificial intelligence. They had a few rules:
1. You can ask about his history, examine him, or order an investigation.
2. You can only make one diagnosis, once.
3. And everything you order goes on a bill that gets added up at the end.
The study used two case repositories: NEJM (New England Journal of Medicine) case records, and MIMIC-IV-ED, a database of real emergency department visits.
In both of them, the AI Agents beat the medical doctors on accuracy. On the hard NEJM cases it got 83%. The doctors got 22% from their own heads, and 36% when they were allowed textbooks and the internet. On ordinary casualty presentations it was 80% against 47%. And the Agent spent less doing it.
How The Models Work:
It is not really a model like ChatGPT or Claude. It is called the MAI Diagnostic Orchestrator. Think of it as a set of instructions for the model to follow.
It takes one model and makes it act like a council of five doctors who argue before anything gets ordered.
1. Dr Differentials, called Dr Hypothesis in the paper, keeps a top three differential and updates it as results come in.
2. Dr Test-Chooser picks the tests and investigations that separate them.
3. Dr Challenger attacks the leading diagnosis.
4. Dr Cost-cutter, called Dr Stewardship in the paper, asks if something cheaper will do.
5. Dr Checklist makes sure the tests are real.
Basically like an Internal Medicine ward round.
Here is another one from the paper.
A man is in hospital for alcohol withdrawal and becomes intoxicated on the ward.
What investigations do you order?
The base model, (meaning the AI without this set of instructions, without the MAI Diagnostic Orchestrator), said antibiotic toxicity. It ordered a brain MRI and an EEG, got the diagnosis wrong, and spent $3,431 of the patient's money on tests.
On the other hand, using MAI, the council asked him questions first. Turns out he had been drinking hand sanitizer. One toxicology investigation confirmed it, and the investigations were significantly cheaper (~700)
They tried the same setup on other models too, o3, Claude, Gemini, Grok, Llama and DeepSeek. Nearly all of them got better results, by about eleven points on average, and the ones that were already strong got cheaper instead. Then they tried it again on newer models without changing anything about the setup, and it still worked.
Limitations of the study:
1. The doctors sat the test alone. Most of them had no books, no colleague to ask, and no chance to go back and examine the patient again. That is not how any of us work.
2. There were no healthy patients in the cases, and no mild illness. Everybody in there had something. So nobody knows how often the AI would have worked up a patient who was fine.
3. There is no real patient here. The case is an old published report, and a model is reading it out to you. Now say you ask for a blood film and that report never mentioned one. The model does not tell you it has no answer. It writes a result for you that fits the case, and you cannot tell the difference. So in this test, nothing is ever unavailable.
How this changes the World We Work In:
First, significantly cheaper healthcare. Models working alongside medical doctors would order better and cheaper investigations for their patients.
Secondly, better diagnosis. Human doctors are brilliant, but when I am 24 hours into a 36 hour shift, my brain slows down. When I am that tired, having a smart partner is great.
A 52-year-old man is admitted to a hospital with fever, nausea and breathlessness that has worsened until he needs oxygen.
What investigations do you order next?
Researchers at Microsoft did a study comparing physicians with artificial intelligence. They had a few rules:
1. You can ask about his history, examine him, or order an investigation.
2. You can only make one diagnosis, once.
3. And everything you order goes on a bill that gets added up at the end.
The study used two case repositories: NEJM (New England Journal of Medicine) case records, and MIMIC-IV-ED, a database of real emergency department visits.
In both of them, the AI Agents beat the medical doctors on accuracy. On the hard NEJM cases it got 83%. The doctors got 22% from their own heads, and 36% when they were allowed textbooks and the internet. On ordinary casualty presentations it was 80% against 47%. And the Agent spent less doing it.
How The Models Work:
It is not really a model like ChatGPT or Claude. It is called the MAI Diagnostic Orchestrator. Think of it as a set of instructions for the model to follow.
It takes one model and makes it act like a council of five doctors who argue before anything gets ordered.
1. Dr Differentials, called Dr Hypothesis in the paper, keeps a top three differential and updates it as results come in.
2. Dr Test-Chooser picks the tests and investigations that separate them.
3. Dr Challenger attacks the leading diagnosis.
4. Dr Cost-cutter, called Dr Stewardship in the paper, asks if something cheaper will do.
5. Dr Checklist makes sure the tests are real.
Basically like an Internal Medicine ward round.
Here is another one from the paper.
A man is in hospital for alcohol withdrawal and becomes intoxicated on the ward.
What investigations do you order?
The base model, (meaning the AI without this set of instructions, without the MAI Diagnostic Orchestrator), said antibiotic toxicity. It ordered a brain MRI and an EEG, got the diagnosis wrong, and spent $3,431 of the patient's money on tests.
On the other hand, using MAI, the council asked him questions first. Turns out he had been drinking hand sanitizer. One toxicology investigation confirmed it, and the investigations were significantly cheaper (~700)
They tried the same setup on other models too, o3, Claude, Gemini, Grok, Llama and DeepSeek. Nearly all of them got better results, by about eleven points on average, and the ones that were already strong got cheaper instead. Then they tried it again on newer models without changing anything about the setup, and it still worked.
Limitations of the study:
1. The doctors sat the test alone. Most of them had no books, no colleague to ask, and no chance to go back and examine the patient again. That is not how any of us work.
2. There were no healthy patients in the cases, and no mild illness. Everybody in there had something. So nobody knows how often the AI would have worked up a patient who was fine.
3. There is no real patient here. The case is an old published report, and a model is reading it out to you. Now say you ask for a blood film and that report never mentioned one. The model does not tell you it has no answer. It writes a result for you that fits the case, and you cannot tell the difference. So in this test, nothing is ever unavailable.
How this changes the World We Work In:
First, significantly cheaper healthcare. Models working alongside medical doctors would order better and cheaper investigations for their patients.
Secondly, better diagnosis. Human doctors are brilliant, but when I am 24 hours into a 36 hour shift, my brain slows down. When I am that tired, having a smart partner is great.
From an ML perspective there is a lot that could still be better here.
Imagine models fine-tuned specially with the MAI Diagnostic Orchestrator in mind, And they open sourced the benchmark for a set of the NEJM cases, the gatekeeper, the judge and the cost estimation with it.
How would this work in Nigeria and Africa?
Most hospitals are adopting EHRs (Electronic Health Records). A model running offline, built for this purpose would provide cheaper investigations, faster diagnoses and improved quality of care for patients.
I'm excited about the future of AI in healthcare. I just think the version that works here is going to have to be built by someone who works here, a Clinician.
