# Mastering LLM Deployment in Healthcare: Strategic Insights From the Industry _Date: 26-03-2024 , Tags:_ #data-science #LLM #GenerativeAI #AI #ML #healthcare #digital-health #corporate ![[adryhealtharticles.png|200]] ## Introduction Deploying LLM models in healthcare requires regulatory compliance, strategic planning, and productive partnerships. How is this achieved in practical terms? Today, I had the privilege to sit down and discuss these topics with Neel Das, LLM Tech Lead @Roche. Neel is an AI leader who has taken an unconventional route in the healthcare industry. With over a decade of experience in machine learning, he now plays a key role at Roche in developing Gen AI solutions for their digital healthcare software offerings. Neel's position is pivotal in generating evidence for the safe deployment of AI models and fostering effective collaboration among data, engineering, and product teams. The views expressed by Neel in this article are his own opinions and observations. ![[Mastering LLM Deployment in Healthcare.png]] ## Hello Neel, we appreciate your participation in the Digital Medicine Beats Newsletter. Could you tell us more about your professional path at the intersection of AI and Biomedical Sciences? _My background is quite diverse. I initially pursued a degree in ocean engineering during my undergraduate studies, followed by three years as a software developer. Afterward, I got a master's degree in a related field._ _During my time pursuing my master's degree, I became increasingly intrigued by machine learning, particularly deep learning, which was just beginning to gain momentum. During that period I was also seeking a career transition to a field where I could make a meaningful impact and find fulfillment in my work._ _Fortunately, I stumbled upon a PhD position for students with strong mathematical and ML expertise. The topic revolved around diagnostic AI models for healthcare. I researched diagnostic AI models and had the chance to collaborate on a startup initiated by my promoters, which eventually produced software products utilized by pharmaceutical companies._ _After nearly seven years in academia, I transitioned to Roche in 2022 as a senior data scientist, where I now collaborate with various product teams._ _Currently, I lead a team of NLP engineers focusing on developing LLM-based features for our external-facing products used by healthcare professionals and patients._ ## How do you shape the product strategy of a new potential LLM-based solution _While I am not directly involved in defining the product strategy, which is typically the responsibility of the product manager and business team. However, with the advent of Gen AI, it has become more important than before for the product teams to work closely with data scientists like us._ _The first step typically involves product managers identifying pain points faced by users that could be addressed through LLM features._ _Subsequently, product managers collaborate with data scientists to develop features that effectively address these pain points, prioritizing initiatives based on business value and potential ROI. Product teams must work closely with data scientists to manage expectations and ensure the feasibility of proposed features, considering limitations imposed by current technological capabilities._ _So, before the PoC phase, the product and data teams build business cases to help this process._ ## In your experience, what are the most common evidence-generation flaws made by data scientists when deploying an LLM into production? _Although there is no playbook, innovators should consider all evidence that supports safe deployment._ _Indeed, the level of evidence required depends on how the AI model is classified by regulatory bodies. For instance, if the model is considered a medical device, extensive scientific evidence, such as clinical trials, is necessary to demonstrate safety and efficacy. This process cannot be expedited, as different classes of medical devices have varying evidence requirements._ _On the other hand, for AI products not classified as medical devices or class I devices, we need to demonstrate that these technologies are still safe and accurate. In the case of LLMs, internal validation studies are essential to demonstrate limited to no hallucinations. These studies are particularly relevant, especially in scenarios where the model's output is inputted into another system._ _Specific challenges related to LLM evaluation include also the complexity of validating text outputs. Unlike traditional machine learning predictions, which can be validated quantitatively, text outputs require more nuanced validation methods._ - _Common approaches to building a hallucination framework involve assessing text eligibility using various criteria, though this can be challenging. Strategies like quantifying similarity in summarization tasks offer partial solutions but may not capture the full context of the text._ - _To address these challenges, a two-pronged validation approach combining automated machine learning techniques with human validation has been effective in ensuring accuracy at scale. Continuous improvement in validation methods is sought to enhance the validation process further._ _Furthermore, the abundance of intellectual property surrounding LLMs has led to varied expectations from executives, requiring effective communication and expectation management between Product Managers, Data Scientists, and stakeholders._ ## Do you have any best practices for collaboration between the data, engineering, and product teams? Is there something that can be scaled across other organizations? _From my experience, establishing best practices for effective teamwork can be challenging due to varying levels of understanding and capabilities across teams. There is no one-size-fits-all approach to collaboration in this dynamic environment._ _There is a role called Data lead who often find themselves navigating between different responsibilities, including data science, strategy development, and stakeholder engagement. Flexibility and adaptability are key attributes for individuals in these positions as they must be able to switch tasks and responsibilities seamlessly to meet the demands of the project_ _While there isn't a singular solution that can be universally applied to all organizations, the ability to be versatile, pivot between different roles, and adapt to changing circumstances is crucial for fostering collaboration and driving innovation within cross-functional teams._ _Indeed, a data leader is a multifaceted role that involves not only managing datasets but also engaging with various stakeholders to address data privacy concerns and leverage advanced data strategies to gain a competitive edge in the market._ ## What counterintuitive advice would you give to yourself, looking back at your journey? _Reflecting on my journey, one counterintuitive piece of advice I would offer myself is the importance of balancing technical skills with soft skills. Initially, I prioritized enhancing my technical expertise in areas like data science and machine learning._ _While this focus was valuable, in hindsight, I recognize the significance of cultivating strong people skills and communication abilities._ _There were instances where I had to forego certain opportunities that required adept interpersonal and communication skills, areas I had not emphasized due to my singular focus on technical proficiency. However, as my career progressed, I found myself gravitating towards roles that necessitated strong communication and relationship-building capabilities._ ## Is there any point that you think would be interesting for data scientists, product managers, or strategic leaders to know (e.g. recommendations could include research papers, a book, or suggestions from your experience)? _I would like to recommend a book that many of you may have already read, "The Lean Startup." This book resonated with me due to its approach to product development, which emphasizes starting with a minimally viable product and iterating based on customer feedback._ _Coming from an academic background, I appreciated the author's insights on the iterative nature of product development and the value of quick feedback loops in improving products efficiently._ _The concept of treating product development as a series of experiments, as highlighted in "The Lean Startup," is applicable not only in building products but also in research._ ## Before we wrap up, is there anything specific you would like to ask the community? _I'm eager to hear about the readers' experiences with how leaders at their companies perceive innovation from frontline workers and how they have successfully scaled up projects based on these experiences._ ## Key Takeaways: - **On Shaping a Product Strategy for LLM-Based Solutions**: Product Data Manager should identify user pain points as a first step, and early involve data scientists to ensure the proposed solutions are double and impactful. - **Evidence level needed to Deploy Medical LLMs**: Varying levels of evidence are required based on regulatory classification. When it comes to LLMs a two-pronged validation approach combining automated machine learning techniques with human validation has been effective in ensuring accuracy at scale. Intellectual property rights considerations cannot be overlooked as well - **The Job of a Data leader**: requires flexibility, adaptability, and the ability to pivot between roles, and establish best practices for effective teamwork. - **On iterative testing (The Lean Startup)**: Starting with a minimal viable product and quickly iterating based on customer feedback, is a best practice that can be applied both in product development and academic research. - **On Career Advice**: soft skills are as important as hard skills. Strong communication and relationship-building can propel professional growth. - **A question for the community**: What are the readers' experiences with how leaders at their companies perceive innovation from frontline workers? How do they have successfully scaled up projects based on these experiences? --- _This article is a repost of the original Digital Medicine Beats Newsletter. Subscribe to Digital Medicine Bests to receive new content._ <iframe src="https://digitalmedicinebeats.substack.com/embed" style="min-width:100%;height:300px"></iframe> --- _Copyright © 2024 - Adriano Fontanari_