AI for Healthcare: Pt 3, Applications
Today AI offers better analytics, personalized medicine, and chatbots to support healthcare. Soon, AI-enabled medical technology acceleration will deliver radically improved healthcare outcomes.
Preface
Note on terminology: Medicine is the diagnosis, treatment, and prevention of diseases, both the practice and the profession. The term ‘healthcare’ can encompass medicine as well as the practice of maintaining and improving physical and mental well-being. The healthcare term also describes the industry to provide these services. As such, there are software and AI applications in healthcare (in administration, insurance, prevention, and wellness) that fall outside the stricter scope of medicine.
Healthcare Algorithms, LLMs and Applications
In our prior article, AI for Healthcare: Pt 1, The Rise of AI, we noted pre-LLM uses of AI in healthcare, pointing out three types of algorithms: NLP (Natural Language Processing), image recognition with CNNs, and RL (Reinforcement Learning). LLMs and generative AI have now augmented “traditional” machine-learning models with additional categories of AI systems in healthcare.
Our second installment in the series, AI for Healthcare: Pt 2, Medical LLMs & AI Research, looked at medical LLMs and the research and development of that technology to serve in many healthcare applications.
LLMs constitute the most significant type of AI model for medical uses, and LLMs and other generative AI models can support shows a broad range of healthcare applications: Diagnostics (text and imaging AI), patient communication (AI chatbots), EMR handling (automated ICD coding, summarization), drug design, visual QA on medical images, and others.
However, we left with a note that:
The impressive advancements of foundation models have not yet permeated into medical AI.
For all its dependence on advanced technology, healthcare is (for legitimate reasons) a risk-averse industry, slow to adopt recent technologies, practices, and innovations, with regulatory approvals and a culture of caution in the healthcare industry.
AI Adoption in Healthcare
We gave a generalized view of AI adoption in the previous article “AI Adoption: From Hype to Utility.” The adoption cycle of AI for healthcare organizations follows the trajectory we described for business organizations in general: Evaluate, Pilot, Adopt, Optimize, Standardize.
LLMs and generative AI applications for healthcare are mostly still in evaluations and pilots. Real-world use adoption is ramping up, and more mature AI methods, such as machine-learning classification models, are already in the adoption phase. We can expect waves of adoption of new capabilities.
To see the potential of AI in healthcare, consider medicine as a special type of information technology. To understand the state of someone’s health, you need to collect a lot of specific data:
Blood and other biomarker tests, physical examinations, blood pressure, heart rate / EKG HRV, MRI, urine, symptom measurements both observed and self-reported.
Understanding of the patient through the patient history, genetics.
To make an informed diagnosis, you match those symptoms, bio-metric measurements, and patient history with conditions. Then you develop a treatment plan, implement those steps, and communicate with patients.
Much medical diagnostics has been expert pattern-matching and reasoning over biological data. AI should not only be able to do this task well, but enable a more nuanced appreciation of diagnostics, because machine intelligence is better at handling vast amounts of data compared to humans. It’s all about mapping data and properly interpreting complex information.
It also shows the major challenge for AI in medical applications: Creative answers aren’t desired, and hallucinations can be fatal. When it comes to making major decisions with reliable accuracy, AI isn’t ready.
As this article “Fact-checking the new o1 'Strawberry' for healthcare” makes clear in it’s review of o1 for healthcare, “even the best large language models (LLMs) we have are not ready for prime time in healthcare.”
Just as automated self-driving cars have been ‘on the horizon’ for over a decade, AI in healthcare has been in a ‘trial phase’ but not yet ready to automate medical processes.
AI in Healthcare – Today’s Applications
Given AI’s current capabilities, how can AI add value to healthcare right now? The answer is to put AI in roles that are not making conclusive decisions but instead managing complex data to support doctors and healthcare providers and automating some patient interactions and administrative tasks. AI can automate support systems that improve analysis and processes while letting doctors make final decisions.
Some examples of where AI is used right now in products and applications, in that supporting-not-deciding role:
Operational Efficiency: AI streamlines administrative tasks, saving time and resources. Oscar Health is using LLMs to improve clinical documentation and process claims more efficiently:
With OpenAI’s API, we have cut the time spent documenting medical care conversations and reviewing lab test results by nearly 40%, saving countless hours and less tangibly, reducing provider burnout by allowing nurses and clinicians to focus on higher-order tasks.
The claims assistant has reduced the time it takes for the claims processing team to resolve escalations by 50%, with accuracy on par or better than human agents. We expect to automate investigation for at least 4,000 tickets per month, or 48,000 tickets by the end of the year.
Personalized Medicine: AI-supported digital twin technology is a great enabler for personalized medicine, as it can represent the unique characteristics of each patient (genetics, health history, conditions) and design appropriate treatment accordingly. Siemens Healthineers has done this with Sophie, a digital twin technology for cancer:
All medical data collected over the course of Sophia's treatments flow into the digital model of the patient: her digital patient twin. AI continuously evaluates her data in real time so that it can make predictions about the development of Sophia's health. Large language models help to structure disordered data and convert this information into medical reports in language appropriate for Sophia's family doctor, her oncologist, or for Sophia herself.
AI is enabling personalized medicine in other ways. Major cancer care institutions have formed the Cancer AI Alliance (CAIA) with $40 million in funding to advance precision medicine through AI. They will share data securely among organizations while complying with HIPAA regulations, with a goal of transforming cancer treatment approaches.
Diagnostic Analysis: While general LLMs cannot be trusted, AI platforms can streamline data collection and analysis for specific disease conditions and improve patient outcomes. For example, AIdocs was able to analyze PE (Pulmonary Embolism) patients at Mount Sinai and assess their risks (and appropriate treatment) more accurately and quickly:
“Speed is everything when managing conditions like PE,” Dr. Dadrass emphasized. “For a patient, the difference between waiting hours or even days for a diagnosis and receiving immediate care can be the difference between life and death. AI accelerates this process significantly.”
AI Chatbots for Healthcare: The Medical Futurist mentions10 AI chatbots, including OneRemission, Youper, Ada, and virtual health assistant Florence, and there are many more examples that have been released. These AI systems improve patient engagement with medication reminders, answering health inquiries, sharing health information, and providing preliminary diagnostics.
On the Bleeding Edge
AI is rapidly improving on medical tasks, and as it does, it is expanding into higher-value applications in healthcare.
As advanced LLMs increase reasoning, knowledge retrieval and textual understanding, they will become essential copilots for clinicians. The health record will become the AI-enabled digital twin, and AI systems will interpret and communicate patient status, conditions, and treatments.
AI algorithms enhance image interpretation in radiology, providing more accurate readings and supporting radiologists in diagnosing from X-rays, CT scans, and MRI images. As AI models go multi-modal, AI will interpret images, text, and other biometrics for the whole human body to yield a holistic diagnostic result.
AI for Robot-Assisted Surgery: Robots, powered by AI, assist surgeons with precision and control, reducing recovery time. Better AI will further automate and improve this process.
Our Healthy Future with AI
The CEO of Anthropic Dario Amedio wrote a lengthy discourse on how AI Could Transform the World for the Better called “Machines of Loving Grace.” In it, he speculates on AI’s potential to accelerate biological science understanding and improve medicine.
He points out that intelligence and its beneficial outcomes can be constrained by data, physical limitations, constraints from humans. All of these come into play in the fields of biology and medicine: Complexity abounds in living systems, data is hard to get, you need real clinical trials, and social and Government constraints will slow progress for safety and other concerns.
Still, he insists that “AI can truly accelerate biology … [by] using AI to perform, direct, and improve upon everything biologists do. … ”
The success of AlphaFold/AlphaProteo at solving important problems much more effectively than humans … should point the way forward. … Thus, it’s my guess that powerful AI could at least 10x the rate of these [biology] discoveries, giving us the next 50-100 years of biological progress in 5-10 years.
AI’s acceleration potential can be found in many areas.
AI-accelerated simulations of protein interaction modeling can help improve the likelihood of success. In Phase I trials, AI-discovered molecules are substantially more successful than historic industry averages.
AI can identify many more drug candidates at far lower costs, widening the net of discoveries.
It won’t replace live clinical trials, but it can accelerate them.
AI can assist in personalized medicine: “AI will also make possible treatment regimens very finely adapted to the individualized genome of the cancer”
Amedio goes further than simply promising AI-accelerated medical advances. He predicts that it will produce a century’s worth of medical advances in the next 10 years that will radically improve lifespans and end most diseases. Within 7-12 years “if aggressive AI timelines prove correct” he predicts:
Prevention and treatment of most infectious diseases
Elimination of most cancers (95%+ reduction in mortality and incidence)
Prevention and cures for most genetic diseases
Prevention of Alzheimer's disease
Improved treatments for ailments like diabetes, obesity, and heart disease
Greater biological freedom, allowing control over one’s weight, appearance, and reproduction.
Doubling of the human lifespan to around 150 years
He adds:
This would represent an unprecedented humanitarian triumph, eliminating many of the health issues that have plagued humanity for millennia.
I am optimistic that AI can accelerate biological science progress. I am skeptical it will translate into curing deadly diseases or extending lifespans as soon as he predicts. We face many hurdles in taking new medical technologies from the lab to the real-world.
Still, this is a great mission to aim AI towards: Use AI to advance biological science, so we can cure major causes of death and double healthy lifespans. Helping us all live longer, healthier lives could be AI’s greatest benefit to humanity.