Evolving Artificial Intelligence Paradigms in HealthCare

Remember that artificial intelligence (AI) is still in its early stages, comparable to the first automobiles in 1900—quite different from today's electric vehicles. It's apparent that the structures of generative AI models will inevitably evolve. The current trend of creating larger, more resource-intensive models requiring vast amounts of training data isn't sustainable in terms of GPU usage and electricity consumption.

We're beginning to see new, more specialized models emerge. The future of AI is increasingly focusing on AI agents, autonomous systems capable of understanding, learning, and solving real-world problems across various domains. This goes beyond traditional language models to more dynamic, goal-oriented, and interactive technologies with the potential to significantly alter how we work and interact with digital environments.

When combined with federated machine learning (FL), AI agents show great potential for collaboration, enhancing privacy-preserving and decentralized learning and action-taking capabilities across various domains. Designed to perform autonomous tasks and make decisions, AI agents could use FL to train on diverse datasets distributed across multiple locations without compromising data privacy. This would allow AI agents to access a wider range of knowledge and skills while complying with privacy and security regulations, especially in sensitive fields like healthcare.

By merging these technologies, we could develop more robust, versatile agents capable of learning from decentralized data sources and applying this knowledge in real-world applications. This would improve both the adaptability and ethical standards of AI deployments. While it may seem like science fiction, with many foundational elements still being established - like interoperability, standards, efficiency of learning algorithms when dealing with decentralized data, new security threat- this is indeed a potential path forward.

Advancing Healthcare through Strategic Data Sharing: The Role of Federated Learning in Collaborative Innovation

According to BCG, the Benefits of Data Sharing Now Outweigh the Risks, but many companies remain hesitant to engage in strategic data sharing. However, successful examples exist in other industry’s, such the Airbus's Skywise platform, which demonstrate substantial benefits from collaborative data utilization. Technical solutions exist, such as AWS Clean Rooms providing a secure environment where data from multiple parties can be analyzed without the parties ever having direct access to each other's data. We see also specialized platforms such as Decentriq or Seqster ’s data sharing platform used by The Novartis-backed Multiple Sclerosis Implementation Network (MSIN). Similarly, Loren Frank's HHMI lab at UCSF has launched "Spyglass," a groundbreaking data analysis framework. This framework facilitates reproducible and shareable neuroscience research, aiming to foster collaboration over competition in the scientific community by standardizing and sharing vast datasets and analysis methods.

Another way to share data is to use Federated learning, a decentralized machine learning method introduced in 2017,which emerged as a pivotal innovation in the life sciences. This technique enables collaborative model training across multiple devices while maintaining data localization, which ensures enhanced privacy and efficient bandwidth usage by sharing only model updates. In medical settings, such as the detection of kidney disorders, federated learning combines with transfer learning to analyze imaging data from diverse healthcare providers without sharing the underlying patient data. This approach not only secures patient privacy but also enhances diagnostic accuracy across decentralized systems. Furthermore, projects like the "Better" initiative demonstrate federated learning's potential in genetic disease research by developing predictive models collaboratively across European hospitals, while adhering to strict data protection laws. Despite its promise, the adoption of federated learning in healthcare faces challenges, with most studies being proof-of-concept, and only 5.2% of initiatives applied in real-world settings. Challenges include the synchronization of updates across datasets and the need for more substantial clinical validation. However, with ongoing research and initiatives, federated learning continues to show potential for transforming data-driven healthcare by facilitating secure, collaborative advancements in medical research and patient care. A LinkedIn training on federated learning covers privacy and intellectual property in cross-organizational machine learning setups. It demonstrates a global federated learning setup using Azure ML and NVIDIA Flare with a Kaggle chest X-ray dataset, simulating local models in hospitals worldwide managed by a global model on a federated server.

Generative AI in Healthcare: Balancing Innovation with Sustainability as LLMs Approach Resource Limits

Are LLMs About To Hit A Wall? I am frequently asked about the sustainability of LLM, considering that the energy demands for running advanced LLMs are becoming a critical constraint, with discussions around the need for gigawatt-scale data centers to support future AI developments. This raises questions about the economic and environmental viability of continuing to expand LLM capabilities at the current rate. If we consider that Meta, for instance, trained its new Llama 3 models with about 10 times more data and 100 times more compute than Llama 2. Amid a chip shortage, it used two 24,000 GPU clusters, with each chip running around the price of a luxury car.

”It is unclear whether we need to continue scaling or whether we need more innovation on post-training,” Ahmad Al-Dahle , Meta’s VP of GenAI, said in an interview last week. “Is the infrastructure investment unsustainable over the long run? I don’t think we know.”

As the industry approaches the upper bounds of data, compute and energy resources, there is a shift towards innovating more efficient model architectures and training methods. These innovations include using synthetic data for training and developing custom hardware like AI-specific chips to enhance the efficiency and performance of LLMs. For more on this topic, listen at Patel Dwarkesh webcast with @Mark Zuckeberg. This trend is confirmed by a survey from JohnSnowLabsGenerative AI in Healthcare Survey which reveals a strong preference for custom-built, task-specific language models.

GeanAI enhances as well real-world evidence (RWE) by efficiently organizing and analyzing patient data, and improving diagnostic accuracy through advanced processing of medical records and imaging. It also bridges data gaps to ensure comprehensive data integration and privacy. However, if Large language models (LLMs) have attracted significant interest for automated clinical coding and early data shows that LLMs are highly error-prone when mapping medical codes: GPT-4 demonstrated the best performance, with the highest exact match rates for ICD-9-CM (45.9 percent), ICD-10-CM (33.9 percent), and CPT codes (49.8 percent). If we can expect further improvements to be done with LLM usage for clinical data mapping, the results are not different for translating free-text into structured reports: OpenAI's ChatGPT models, GPT-3.5 and GPT-4, are making strides in radiology by translating free-text thyroid ultrasound reports into structured reports adhering to ACR-TIRADS guidelines; but more consistency and uniformity across all tasks is required to match a clinical grade.

Additional news on Generative AI:

• What We Know About Apple's Work on Generative AI? Apple, which is seen as being late on generative AI, has been testing an "Apple GPT" rival that could compete with ChatGPT. According to Bloomberg's Mark Gurman, work on AI is a priority for Apple, with the company designing an "Ajax" framework for large language models. Apple is rumored to be planning to add AI to as many Apple apps as possible. We should also see more "On-Device AI" with the first AI features coming to the iPhone with iOS 18 running entirely on-device using the iPhone's hardware rather than relying on a cloud server. See what you can expect.
• Moderna uses AI at scale to reinvent all business processes: Moderna's collaboration with OpenAI, focusing on the integration of the AI company's advanced language models (specifically GPT-4), illustrates a significant shift in how biotechnology firms are leveraging AI to enhance operational efficiency and accelerate product development. Weeks to minutes in some processes.
• Snowflake introduces Arctic LLM, an enterprise-grade generative AI model optimized for business applications like SQL co-pilots and chatbots. This reflects a shift in cloud vendors focusing on enterprise needs with specialized, cost-effective models that include MoE (Mixture of Experts) architecture to enhance efficiency.
• Gen AI and infinite context: Google researchers have developed a new technology called Infini-attention that allows language models to process very long texts without forgetting earlier parts. Traditional models lose previous information when handling new data, but Infini-attention keeps and compresses all past data, making it accessible at any point. This method enhances performance, allowing models to handle up to 1 million tokens effectively and achieve top results in summarizing long documents like books. Infini-attention also maintains consistent memory use and reduces computational demands, making it efficient and scalable for large-scale tasks.

Drug Discovery: 3 AI-Driven Biotechs Aim to Surpass Legacy Innovations

The new leaders of the AI biotech field —?Generate:Biomedicines,?Isomorphic Labs?and?Xaira Therapeutics?— are hoping they can do what an earlier crop of equally-hyped companies that launched a decade ago couldn’t. On April 23rd,?the $1 billion-plus launch of Xaira Therapeutics?— the second-biggest initial biotech startup funding ever — took a field that’s already described by some as overhyped and whipped it into a frenzy.?In a great Endpoints News interview,?Alexandra Snyder, MD, head of R&D at Generate:Biomedicines, acknowledges the current excitement in AI-driven biotech as potentially overhyped but significant, mirroring a frothy drink where the true substance underneath is yet to be fully realized. The biotech industry is optimistic about leveraging recent AI advancements (see the new method called a?diffusion model used for creating proteins from scratch) to overcome the shortcomings of previous companies in transforming drug discovery and development.

On geological time, we’re pretty early. The people who made the first airplanes certainly couldn’t imagine a Concorde, even though they had the right idea. Alexandra Snyder

Stifel conducted an insightful report, interviewing eighteen AI experts and their applications in the pharmaceutical industry. The experts included seven individuals from AI biotechs, three investors, three technologists, a few service providers, and six representatives from big pharma. The application of AI in pharma presents dozens of use cases, ranging from drug discovery to technical writing. However, the discussions primarily focused on the use cases in drug discovery and development. Key insights derived from these expert interviews included:

Meanwhile, Insilico Medicine has used generative AI to discover a new class of inhibitors for BRCA-deficient cancers, which show potential in early tests to provide effective treatment options for these resistant cancers. In the Pharma industry, Pfizer's three-year collaboration with CeMM has resulted in an AI-driven drug discovery method. This method rapidly identifies the therapeutic potential of small molecules by mapping their interactions with human proteins. The resulting data and AI models are freely accessible via a web application.

AI Revolutionizes Clinical Trials: Market to Reach $6.55 Billion by 2030, Enhancing Efficiency and Patient Recruitment

The AI in Clinical Trials Market was valued at 1.59 Billion USD in 2023 and projected to reach 6.55 Billion USD by 2030, growing at a CAGR of 22.4% during the forecast period of 2023-2030. Using AI to improve Patient recruitment is one of the drivers. Patient recruitment impacts not only trial timelines but also the overall success of clinical research, and this is one of the main challenges for Life Science company’s. To address this, The National Institutes of Health's National Library of Medicine is exploring the use of GenAI to enhance patient matching for clinical trials. This initiative is part of a broader strategy to leverage digital technologies to overcome the traditional barriers in medical research, particularly in the recruitment of suitable clinical trial participants. A notable innovation from this initiative is "TrialGPT," an open source prototype that uses an LLM framework to predict patient eligibility for clinical trials by analyzing patient notes. This tool aims to streamline the patient selection process by providing detailed explanations of eligibility, thereby improving the efficiency and accuracy of matching patients to relevant trials. However, the initial phases have revealed challenges such as the model's limited medical reasoning capabilities and lack of intrinsic medical knowledge.

Additional recent announcements:

• After partnering with Novartis, which uses Foundry for secondary data usage (data42 initiative), Palantir has now partnered with Parexel. Parexel will leverage Palantir’s Foundry and Artificial Intelligence Platform (AIP) to further enhance its clinical data platform, focusing on improving clinical trial efficiency while maintaining safety and regulatory rigor. Parexel is the first Contract Research Organization (CRO) to work with Palantir in this capacity, expanding on the company’s collaboration over the past year.
• Deciphex, a Dublin-based digital pathology company, has entered into a Collaboration Agreement with Novartis to develop artificial intelligence (AI)-based approaches with the potential to transform preclinical pathology assessments in drug discovery and development. The collaboration seeks to develop a suite of AI tools that streamline the pathology evaluation of toxicological and efficacy studies.
• Crucial Data Solutions (CDS), a leading provider of data collection and trial management software to advance clinical research, today announced the launch of TrialKit PACS. This innovative solution streamlines medical image management within clinical trials, offering a unified platform for all data collection and management needs.
• Nextrials, Inc. has updated its Prism? platform to include a real-time ad-hoc reporting feature, allowing biopharma and medical device researchers to access and analyze data instantaneously for more flexible and faster decision-making.
• Risklick, a spin-off from the University of?Bern, launched Protocol AI, the first software of its kind, marking a turning point in the pharmaceutical industry. By analyzing existing clinical data using Natural Language Processing (NLP) and Large Language Models (LLM),?Protocol AI?facilitates evidence-based decision-making and automatic drafting of clinical trial protocols within minutes.

AI in Healthcare

Increase budget for GeanAI and Shift Towards Customized AI Models for Improved Patient Care

A recent study published in the Lancet Digital Health, suggests that AI has shown potential benefits in healthcare, particularly in improving diagnostic effectiveness and clinical management and often outperforming or matching clinician performance in controlled trial settings. Another study at Brigham and Women's Hospital examined the use of Large Language Models (LLMs) in drafting responses to electronic patient messages. While it found that LLMs improved physician efficiency and response consistency, concerns were raised about potential automation bias and the accuracy of clinical decisions. These findings indicate a need for careful evaluation and integration of LLMs in clinical settings.

In response to this kind of need, Hugging Face, in collaboration with Open Life Science AI and the University of Edinburgh's NLP Group, developed The Open Medical-LLM Leaderboard. It benchmarks the performance of various large language models on medical-related tasks using a collection of medical datasets like MedQA, PubMedQA, and MedMCQA. This initiative aims to evaluate AI models' ability to accurately answer complex medical questions drawn from both U.S. and Indian medical exams and biomedical literature. The goal is to ensure reliability and enhance patient care outcomes by identifying the strengths and weaknesses of different AI approaches in healthcare settings.

AI : Surpassing Human Accuracy and Expanding Healthcare Accessibility

There are significant advancements in AI medical diagnostics, where AI is surpassing human doctors in certain areas. For example, GPT-4 outperformed ophthalmology trainees in diagnostic accuracy and AI has matched or exceeded clinicians in analyzing medical images like mammograms and retina scans. AI is also improving cardiovascular risk assessment by identifying subtle signs in CAT scans. Moreover, AI-driven retinal imaging can potentially predict conditions like heart disease and Alzheimer's. As examples, Advanced AI models are now capable of evaluating cardiovascular risk from routine non-contrast chest CT scans more effectively than traditional radiologist assessments. Mayo Clinic has developed an AI algorithm that detects low ejection fraction, a symptom of potential heart failure, now integrated into Eko Health's smart stethoscopes used by over 500,000 medical professionals worldwide following FDA approval. AI applications also improve the detection and management of diabetes complications by leveraging machine learning, deep learning, and fuzzy cognitive maps to analyze patient data, enhance diagnostic accuracy, and predict disease progression. Moving forward, detecting cancer in minutes will also be possible with just a drop of dried blood. Thanks to a new AI-powered test, using less than 0.05 milliliters of dried blood, can quickly and accurately detect gastric, colorectal, and pancreatic cancers by analyzing blood metabolites as biomarkers, according to preliminary research published in Nature Sustainability.

These developments suggest AI's growing role in enhancing precision, efficiency, and accessibility in healthcare. We are on the way to personalised medicine with AI revolutionizing cancer care by synthesizing medical records, imaging, and genomics into personalized treatment plans. This is enabling precise diagnoses, targeted therapies, and faster clinical trial matching through advanced data analysis and multimodal machine learning. A great example of this progress is the new AI model named PERCEPTION which predicts patient response and resistance to treatment using single-cell transcriptomics of their tumors. This new method can predict how individual tumor cells respond to specific treatments, potentially enhancing treatment effectiveness.

AI in Healthcare: partnerships and Announcements

??Roche has partnered with Prenosis to distribute Prenosis' AI-driven sepsis detection tool, Sepsis Immunoscore, in the U.S. This tool, which analyzes 22 health parameters to predict sepsis risk within 24 hours, has recently received de novo authorization, validating its effectiveness.

??Sectra's digital pathology solution, in partnership with Leica Biosystems, has received FDA clearance to use DICOM images for diagnostics, enhancing standardization in digital pathology and supporting the growth of remote diagnostic practices.

??Parkinson: The MAGELLAN AI platform, utilized by Gain Therapeutics, employs AI and physics-based methods to precisely identify novel drug binding sites, significantly advancing the development of treatments like GT-02287 for Parkinson’s by targeting specific molecular interactions.

??The World Health Organization introduced SARAH, an AI-powered chatbot designed to provide basic health information and help bridge the global healthcare worker shortage. Named “Smart AI Resource Assistant for Health”, SARAH can discuss various health topics in eight languages, but faces challenges with accuracy due to outdated training and AI-generated errors.

??Curio Digital Therapeutics Inc. has received FDA 510(k) clearance for MamaLift Plus?, a digital therapeutic designed for the symptomatic treatment of postpartum depression (PPD). The eight-week therapy, delivered via a mobile app, combines Cognitive Behavioral Therapy and other neurobehavioral interventions. Its effectiveness was validated in the SuMMER study, which demonstrated significant symptom improvement.

Technology & Engineering

Platforms host the apps you've built to run your business! They serve as the interface between your application developers and the cloud infrastructure and services their apps use. Choosing between buying or building a platform can be challenging. CIO Magazine evaluates these two options for you.

How to Implement Knowledge Graphs and Large Language Models (LLMs) Together at the Enterprise Level: Integrating Knowledge Graphs with Large Language Models at the enterprise level merges structured data accuracy with advanced natural language processing, enhancing data accessibility and governance for more efficient business operations.

Build Autonomous AI Agents with Function Calling: Function Calling in large language models (LLMs) such as OpenAI's GPT models and Google’s Gemini API enhances AI agents by enabling them to interact autonomously with external APIs. This makes them capable of performing tasks beyond simple conversation by integrating structured decision-making into their responses. This advanced feature allows AI agents to function as autonomous assistants capable of handling complex tasks like customer service or controlling IoT devices, significantly expanding the utility and application of AI in practical settings.

OpenAi is testing the ability for ChatGPT to remember things you discuss to make future chats more helpful. You’re in control of ChatGPT’s memory. The new "Memory" feature in ChatGPT for Plus users, which enhances personalization by remembering details from user interactions is not yet available in Europe or Korea. As an example, if you frequently discuss project management in your chats, ChatGPT can remember your preferred methods and tools. The next time you ask for advice on managing a project, it can tailor its suggestions based on your past preferences, saving you the time of re-explaining your methods.

If you need more data to test you solution, Gretel provides a Synthetic Data Platform for enterprise developers, designed to address the 'data bottleneck'—a common challenge where AI/ML, research, and development teams struggle to access high-quality data quickly. This bottleneck often arises due to issues related to data availability, quality, or sensitivity, which prevents teams from effectively using the data they need.

Microsoft recently launched VASA-1, an AI model capable of creating realistic talking face videos from a single image and audio clip. This model operates on a diffusion-based architecture to produce holistic facial dynamics and head movements at 512x512 resolution and up to 40 frames per second.

Upcoming 2024 Conferences

• 5th MENA Telehealth & Virtual Care Expo, Riyadh?, 7th-8th My 2024i
• HIMSS24 Europe, Roma, Italy, May29th-31st, 2024
• Abu Dhabi Global Healthcare Week, Abu Dhabi May 13th-15th, 2024
• IFCC WorldLab 2024, Dubai, May 26th-30th
• Ai-Med Global Summit 2024: Florida, USA, May 29-31, 2024
• 82nd FIP Wolrd Congress for Pharmacy and Pharmaceutical Sciences, Cape TownSeptember 1st-4th
• Pharma Days 2024, Geneva, Switzerland, June 5th-6th, 2024
• HLTH Europe, Amsterdam, Netherland, June 17th-20th 2024
• Future Labs Live, Basel, Switzerland, June 25th-27th, 2024
• BiotechX USA 2024, Philadelphia, PA, September 10-11, 2024
• Intelligent Health AI, Basel, Switzerland, September 11-12, 2024
• BiotechX Europe 2024, Basel, Switzerland, October 9-10, 2024
• Med Tech World, MALTA, November 6th, 8th, 2024
• 6th GCC eHealth Workforce Development Conference, Dubai, November 6th-7th, 2024
• DigiHealthDay 2024, Pfarrkirchen, Germany, November 11th-15th
• GIANT Health 2024 , London, December 9th-10th, 2024

Other Great Newsletters on Healthcare and AI

Stay up-to-date with the latest insights and industry trends by subscribing to these informative newsletters from my esteemed colleagues and professionals in the field:

• Healthcare in the Metaverse by Philippe GERWILL
• DigitalhealthReadingList by Wolfgang Schleifer
• Future Tech Trends by Bernard Marr
• All about Blockchain Adoption by Kamlesh Nagware
• AI Frontier by Steve Nouri
• The Atlas by Generative AI
• Last Month in Digital health by Gary Monk

AI Basics

DataCamp:?AI for Beginners

Google’s?AI Video Course

Matt Turck:?The AI/ML Landscape

MIT:?AI Basics

Nvidia: Jensen Huang on AI

SimpliLearn:?AI Bootcamp

WSJ:?AI Predictions and Future of Humanity