An MIT professor gave a programming challenge in Fortran to 3 groups in his class.
No one knew how to use Fortran.
Group 1 used ChatGPT.
Group 2 used Meta’s Code Llama.
Group 3 used Google.
Guess who succeeded?
The ChatGPT group solved it the fastest, while the Google
Google just unveiled its 2024 vision for Generative AI in healthcare, and it aligns with what I have been sharing in my public talks.
Essentially, there are 3 main opportunities:
1️⃣ Near-term: Reducing administrative work
The immediate focus is on reducing the administrative
100k followers << 1k emails
An email list is your direct line to potential customers but building one is tough.
Writing newsletters on top of daily content? Stressful.
Engagement? Doubtful.
There is a better way, Let's Quizz
I am excited to introduce Let’s Quizz which turns
I am an "AI engineer", and I prefer this term over "prompt engineer"
My responsibilities include
- prompt engineering
- evaluate different LLMs
- optimizing inference costs
- use tools like LangChain
- know LLM app architecture e.g RAG
- fine-tune
- keep up with the news
Can we benchmark open source medical LLMs? Thanks to
@huggingface
, we now have the Open Medical-LLM Leaderboard 🚀
This leaderboard tracks and ranks LLMs based on their performance in medical question answering tasks.
It evaluates LLMs on a variety of medical datasets,
Just gave a lecture on the history of clinical NLP at the University of Zurich!
It felt so good to teach and interact with students after a long time 😄
PS: For those who don't know me I work on medical LLMs 😊
Can AI predict a patient's future health events?
Foresight, an AI model that combines text and structured data from EHRs, shows promising results in forecasting medical outcomes like disorders and procedures with high accuracy.
Applications include risk forecasting, virtual
How can we identify biases and inequities in healthcare applications of LLMs?
A recent work from Google presents a comprehensive framework for evaluating and identifying biases that might cause harm and exacerbate health disparities in LLM applications.
Key contributions of the
Google’s recent conversational diagnostic AI system (AMIE) was not trained in the usual way as you’d think. Instead of training on real-world medical conversations, they took a different approach
Google researchers identified two significant issues with real-world medical
I'm excited to kick off the new year with loads of energy! My annual break gave me time to think about what I really wanted to share with you all.
I've decided to focus on topics that align with my expertise, profession, and passion. The answer was clear: AI in healthcare and
Key announcements from Google at HIMSS 2024:
- Vertex AI search for Healthcare for quick parsing of patient data and clinical notes
- Upgraded MedLM, a medical LLM model with new features for x-ray classification and summarizing patient conditions
Startups let's get real! Wrapping ChatGPT API in a UI is not called a business model. It is a cash grab at best.
Good news, you can still create one if you:
- control exclusive datasets
- solve an incredibly tough problem
- have a habit-forming product
#buildinpublic
Are you interested in learning healthcare data science? Here are 5 free courses taught by leading universities in the world.
1️⃣ Clinical Data Science Specialization by University of Colorado ()
This specialization provides an overview of different clinical
Generic advice like "be focused", "wear headphones" etc. do not work when it comes to avoiding distractions while coding.
Surprisingly, nothing works better than using a timer, which invokes a sense of urgency to finish tasks.
The Pomodoro technique is gold.
#buildinginpublic
How can we determine if an LLM is safe for use in a medical setting?
Presently, we know very little about how safe LLMs are in a medical setting, or how to check and improve their safety.
To answer this question, Han et al. have established a definition for medical safety in
“Should I do a PhD to get a job as a data scientist? Maybe I should enroll in one until I find a job.”
I find this perspective somewhat offensive, as it implies that a PhD is just another job-training certificate or a visa-extension tactic.
Instead, a PhD is an opportunity to
Working on a side project is hard esp when you have a demanding job. These days I am feeling totally burned out. Anyone been in a situation like this? How have you managed this?
#buildinpublic
#indiehackers
AI in drug discovery is one of the fastest-growing yet highly specialized sectors in healthcare. You can now enhance your skills with these online courses and seminars.
1️⃣ Disease Modeling and Target Discovery by Insilico Medicine
Led by the renowned Alexander Zhavoronkov and
@thejustinwelsh
Absolutely! Being underestimated allows us to exceed expectations and prove our true potential. It's a powerful motivator to keep pushing forward and achieve great things.
I am looking for interns who are passionate about CS & programming. I need some help on a project. Good stipend guaranteed. Feel free to DM me.
#100DaysOfCode
#buildinginpublic
12 promising hospital tech startups tackling healthcare challenges, from AI-driven patient care to robotic surgery aids. Innovations aim to improve efficiency, safety, and patient outcomes amidst rising pressures on healthcare systems.
10 clinicians rated Almanac, an LLM framework with retrieval capabilities from curated medical resources. It outperformed standard LLMs (ChatGPT-4, Bing, Bard) in factuality, completeness, user preference, and adversarial safety.
Paper:
@bindureddy
Unfortunately, that's not true. Most healthcare companies want to adopt AI but in a way that is safe and demonstrates benefit for patients. You cannot just unleash AI on high-risk human facing applications without considering equitable care and safety
BioMistral is an impressive open-source effort in medical LLMs, but I find it suffers from 3 critical shortcomings that could limit its application.
1️⃣ Unrealistic evaluation
The reported evaluation relies on multiple-choice question-answering. This method hardly captures the
Wow, now you can still speak if you have lost your vocal cords! This thin device that sticks to the neck translates the muscle movements of the larynx into audible speech using ML.
Are you looking for a career in digital health and AI? Whether you are a researcher, medical professional, technical developer, someone from med-tech, or a biomedical engineer, the Medical Futurist (
@Berci
) has you covered
@akshay_pachaar
That's a really good explanation Akshay. I also found StatQuest's PCA videos very helpful in understanding the core concepts, especially eigenvalues/eigenvectors
Biostatistics is the foundation of healthcare data science. And if books are your best friends, here are 6 exceptional reads to get started without paying a single penny.
1️⃣ Introductory Biostatistics by Chap T. Le
This book offers a clear explanation of fundamental
Are you preparing for a biomedical/healthcare data scientist position? Here are 9 great resources to ace your technical interview.
I have grouped these resources into foundational and role-specific topics.
🎯 Foundational
1️⃣ Statistics (): A handout of
On this Diwali, I feel truly blessed to have a partner who is so amazing, and a passion that I get to work on and write about daily.
Happy Diwali to all 🎇.
Love 💖
Neel and Tulika
What is a segmentation foundation model in medicine and why is it a game-changer❓🤔
Medical segmentation is a crucial task involving identifying and marking regions of interest (ROI) such as organs, lesions, and tissues in various medical images.
Traditionally, this task is
There are 4 fundamental pillars to becoming a biomedical or healthcare data scientist. Start learning them today with these free resources.
1️⃣ Biostatistics
You must know biostatistics as the description, analysis, and interpretation of almost any data in life sciences will
You are not an AI company if all you do is an LLM API call with a fancy interface.
You are an AI company if you have novel algorithms and/or proprietary data on top of an LLM API call.
...and also if you solve a genuine problem that people have.
#startups
#buildinpublic
Can AI-enabled ECG reduce all-cause mortality in hospitalized patients?
A multisite clinical trial involving 39 physicians and nearly 16,000 patients found that an AI system analyzing 12-lead ECGs and triggering rapid response alerts significantly reduced 90-day mortality from
Are you interested in learning machine learning for healthcare applications? Here are 6 free courses to get started with right now.
🎯 Foundational: No prior experience with ML is required
1️⃣ Fundamentals of Machine Learning for Healthcare by Stanford University
Part of the
Bioptimus, a French startup just raised a $35m seed round to build Biological Foundational Models (BFMs). But what exactly is a BFM? 🤔
I spent this weekend researching this topic and found very little scientific research that discussed BFMs.
However, one paper that
Are there effective safeguards to prevent LLMs from generating healthcare disinformation?
A new 12 week follow-up study assessed the ability of major LLMs to resist generating false health claims about sunscreen and diet cures for cancer.
While Claude 2 consistently refused
Developing ML models to predict molecular biomarkers directly from H&E-stained whole-slide images is challenging due to a lack of labeled specimens. Or until now!
A team developed a novel method that synthesizes realistic whole-slide images from genetic info, while preserving
Recently, A Gen AI leader from Google remarked that role of AI engineers like us is becoming similar to that of IT.
These roles are shifting towards implementation rather than an traditional ML R&D and deployment
There is some truth in that. Most Gen AI applications, such as
@svpino
With due respect Santiago, I love your opinions. But saying something a Jupyter notebook killer (probably this app was build on top of it) is too much of a stretch. Nothing can beat the flexibility of Jupyter!
This image from 1963 shows the birth of AI in medical imaging 👇.
What you see is not just an X-ray, but a painstaking process of assigning numerical values to the different visual features in the X-ray onto a punch card.
Each hole on this punch card represented a specific
In Oct 2023, The Permanente Medical Group (TPMC) initiated a regional pilot program, deploying ambient AI scribes to 10,000 physicians and staff members.
This cutting-edge technology was introduced to transcribe clinician-patient encounters in real-time, aiming to alleviate the
Academic publishing may well be the greatest scam in the world!
There exists no other market system, in which the seller must pay to sell their products, and must also pay to use those products.
The middleman (aka the publisher) reaps everything, for free.
Thank god I am out
🌟 3 AI Trends in Healthcare to Watch Out for in 2024 🌟
Healthcare is on the brink of a major transformation, thanks to the power of AI. Here are three key trends that are changing the game:
1️⃣ Generative AI Platforms: These platforms are making it easier for patients to
Want to know how the field of ML has evolved in the last decade or so?
The legendary
@JeffDean
hosted an informative seminar on this, where he walks the viewers through the early days of deep learning and NLP to until now.
When I was a teaching instructor, I spent countless hours reviewing and grading students' code submissions. It was painful to review assignments week after week for 50 students.
Back then, there was no solution, but today there is.
I've been working on a Gen AI-based platform
Can AI match medical experts in summarizing clinical texts? A new study evaluates 8 LLMs for tasks like radiology reports & patient dialogues. Results? 81% of AI summaries matched or beat expert quality, hinting at a future where clinicians spend more time on care, less on
Did you know: Despite advances in molecular testing, there is ONLY ONE approved biomarker to detect Pancreatic Ductal Carcinoma (PDAC). PDAC kills 55.5k people in the US alone and has extremely limited in predicting clinical outcomes
Thanks to a multi-omic ML approach that
One unexpected advantage I've noticed from working with LLMs is that it has significantly enhanced our collaboration with subject matter experts (SMEs) such as clinicians.
Earlier, the involvement of SMEs was limited to providing high level direction, clarification and feedback
Can you hustle your way to an MVP when starting a healthcare AI company?
Unlike tech startups, where a Minimum Viable Product (MVP) might be a bare-bones app to solve a specific problem, the bar for an MVP that delivers medical insights (e.g. risk prediction, diagnosis,
Do you know which country and clinical specialty sees the largest number of randomized clinical trials (RCTs) with AI?
A new review reports that between 2018 and 2023, USA (31%) and China (30%) conducted the most no of RCTs with AI.
A significant majority of the trials were
I am really excited, my MVP is ready just a day before Easter. Will soon start sending out invites soon to all those who signed up. I know it will have bugs/limited features but I am 100% committed to working with you and improving it together. 😊
#buildinpublic
5 things one should know before starting a career as a healthcare data scientist.
1️⃣ 𝐃𝐚𝐭𝐚 𝐃𝐢𝐯𝐞𝐫𝐬𝐢𝐭𝐲: The variety and complexity of biomedical data is immense. This includes demographics, vital signs, lab measurements, clinical notes, medical imaging, genomics, and
Could AI reduce physician time on reading and replying to patient messages by drafting the replies for them?
An interesting study found that instead of decreasing, GenAI increased the time physicians spent reading messages by 21.8%.
Why?
The authors discuss that the increase
Interested in graduating as a healthcare data scientist? Here are the top 10 universities with a master's program in healthcare data sciences 🧑🎓
1️⃣ Harvard TH Chan School of Public Health ()
2️⃣ UC San Francisco Department of Epidemiology and Biostatistics
Clinical decision making in oncology is really complex, as one has to compare and refer multiple guidelines which keep getting changed at a rapid rate. Can LLMs make this process easier?
A recent study assessed if GPT-4 could effectively support oncologists by providing
How can we improve fairness and robustness in medical AI across different conditions and patient groups?
A new study leveraged Gen AI to enhance training datasets with synthetic examples, targeting underrepresented conditions.
By integrating diffusion models, the authors
My project LetsQuizz managed to get 50+ signups in today. All I did was make a few tweaks to my landing page.
Lesson: convey your value prop immediately and backup with social proof!
#buildinpublic
Website:
Recently, I had the honor to speak with Richard Stevenson, who is the director of AI at
@BarringtonJames
. We discussed about about the impact of LLMs transforming clinical care, and the wealth of opportunities it presents. Take a listen 😃
Folks who have a very interesting full time job and doing an equally interesting side project, how are you managing both? How do you avoid context switching?
#buildinpublic
Top 5 myths about AI in healthcare:
1/ AI will replace doctors
2/ AI's accuracy is sufficient for clinical impact
3/ More data is required to train a better model
4/ AI in medicine is all about deep learning on imaging data
5/ FDA-clearance of AI software is the highest form
I'm still trying to grasp the true benefit of open-source LLMs. Although they're freely available, they aren't necessarily economically viable. Perhaps they work for big corporations, but for solo developers or startups, the pay-per-use API model is far more appealing than
Steve Jobs was fired from Apple for Macintosh's struggling sales and internal power struggles.
Jerry Yang was asked to step down due to Yahoo!'s declining fortunes.
Steve Ballmer was asked to retire at Microsoft due to missed opportunities in mobile computing & search.
Travis
How can AI speed up the discovery of new drugs, which typically takes 10-15 years and >2.5 billion $ to reach the market? 🤔
To answer this question, let's first understand how a medicinal drug works, using a simple analogy of a lock and a key. The 'lock' represents a biological
Can LLMs be trusted to make clinical decisions autonomously?
In a study using the MIMIC-IV dataset with 2.4k real patient cases across four common abdominal pathologies, LLMs were evaluated for their capability to autonomously provide medical information.
Unlike prior studies,
5 types of people obsessed with gpt-o on X right now
1. The AI newsletter paparazzi
2. The ChatGPT influencers
3. The slightly technical bro who did a random prompt test to conclude gpt-4o is better/worse than [insert your fav model]
4. The "never have I built an AI product
Screening patients for clinical trials is a costly and error-prone process. Can AI improve it?
In a recent study, a team developed a GPT4 based RAG system called RECTIFIER to identify patients from unstructured patient records.
The system worked by asking 13 criteria questions
Can AI cause behavioral change in clinicians?
A study found a significant reduction in nuclear scans (80% -> 58%) when AI was integrated into a hydronephrosis (kidney) clinic while nothing else changed including personnel.
Turns out that through repeated interaction with AI,
From just a single prompt, you can now create whole Leetcode styled quiz with a code editor, compiler and checker
Prompt: "Generate a Python coding quiz for beginners."
Result 👇
#buildinpublic
Are you interested in learning data visualisation for healthcare applications? Here are 5 great courses that you can start learning for free.
1️⃣ Tableau's Health and Healthcare Data Visualization course
It introduces the essentials of visualizing health and healthcare data,
@ykdojo
True! Remote jobs offer the opportunity to tap into a global talent pool and find top-notch candidates. Although local country regulations are may still be a blocker
There's a common misconception that a biomedical degree and experience are a must for becoming a data scientist in healthcare.
Let me categorize the different types of healthcare data scientist roles:
1️⃣ The Applied Scientist: These individuals work in research settings. Such
“Neel, could you help me with this analysis? I would be happy to have you as a co-author.”
Back then, during my Ph.D., whenever I received such a request from a biomedical researcher, I viewed them as easy wins.
All I needed to do was open my laptop, load a nicely prepared
Many cancer patients from North-East India travel to Mumbai seeking better treatment and medical support. However, these patients endure not just physical and emotional pain but also significant financial hardships.
Recognizing the need to help these patients and their families,
Are you a biomedical student or researcher struggling with data analysis due to limited programming skills? 🤔
TechMedBuddy and I are organising a one-of-a-kind workshop, completely FREE, to help you leverage Generative AI in accelerating biomedical data analysis 🚀.
No, this
I am proud to be part of
@Roche
which has taken the lead in developing guidelines promoting the ethical use of AI in healthcare.
The 11 AI Ethics Principles that guide the design, development, deployment, and monitoring of AI solutions include:
1️⃣ Ethical Use: AI should be used
Is Gen AI in healthcare our Hemingway moment? A Hemingway moment is a time when one experiences a profound realisation, similar to the way Hemingway presented his narratives.
In the 90s when many fields embraced rapid digital transformation, healthcare was late to the party.
What are the different career paths for biomedical or healthcare data scientists?
I've been closely tracking the field of biomedical data science, which has seen rapid changes in the last 3-4 years. Today, there's a wide range of exciting career opportunities available. Here’s
5 interview topics you can prepare to get a job as a LLM engineer 👇
1. Transformer architecture & its impact on NLP
2. Types and applications of fine-tuning
3. Weight quantization & optimizing LLM inference
4. Evaluating LLM-generated text
5. Ethical considerations in LLMs
Tip💡: Developing apps with open-source LLMs? Use quantization to reduce costs.
It scales down model weights from 32-bit to 16/8-bit. This slashes memory needs and boosts speed while keeping accuracy more or less intact.
Microsoft just released a totally FREE Gen AI course for beginners. 🎉
Topics:
🤖 AI & LLM Fundamentals
🔍 Model Selection & Comparison
🌐 Ethical AI Usage
✏️ Prompt Engineering Techniques
📲 Building AI Text, Chat & Image Apps
🛠️ Low-Code Solutions
🎨 AI UX Design