TL;DR:
- HARMAN introduces HealthGPT, a private LLM tailored for healthcare, addressing data privacy concerns.
- Utilizing open-source models like Falcon 7B and Llama 2, HARMAN focuses on refining existing technologies rather than building from scratch.
- HealthGPT tackles data scarcity and quality issues in healthcare by leveraging publicly available clinical trial data and implementing bias correction mechanisms.
- To ensure accuracy and reliability, HARMAN employs automated mechanisms, human validation, and a feedback loop to manage hallucinations.
- HealthGPT operates within the user’s Virtual Private Cloud (VPC) for enhanced security and privacy, emphasizing responsible AI practices.
- Proof-of-concept stories showcase HealthGPT’s potential applications in personalized data analysis, medical instrument data insights, and pharmaceutical drug discovery.
- HARMAN aims to expand HealthGPT’s reach beyond healthcare into manufacturing and IT management, aligning with its strategic objectives.
Main AI News:
As of October 2023, HARMAN, under Samsung’s umbrella, ventured into the competitive arena of generative AI within healthcare with HealthGPT – a proprietary Language Model (LLM) based on TII’s Falcon 7B open-source model. The most recent iteration, HealthGPT Chat, is now powered by Llama 2.
The healthcare sector poses inherent complexities for generative AI integration due to the abundance of unstructured, sensitive data. Dr. Jai Ganesh, Chief Product Officer, articulated to AIM in an exclusive interview, “The genesis of HealthGPT as a private LLM arises from corporate apprehensions surrounding data privacy and security, particularly concerning public LLMs which entail transferring sensitive data externally, leading to potential uncertainties regarding data usage.”
Dr. Ganesh elucidated further on the challenges posed by public models, emphasizing their structural limitations, such as token constraints and lack of end-to-end control. This lack of autonomy can significantly impact business operations if issues arise with public models.
Recognizing the gravity of these challenges, HARMAN pivoted towards developing private models, coinciding with the momentous open-sourcing of foundational models around late February last year. This move aligned with the leak of Meta’s LLaMA 1 online, followed by the release of models like Falcon 7B and Llama 2. However, rather than developing an in-house LLM, the company opted to harness the potential of existing open-source models. “The decision not to develop our foundational model stemmed from the exorbitant cost and resource investment estimated at $30 to $40 million. Instead, we chose to leverage existing open-source foundational models,” Dr. Ganesh remarked.
Data Dearth
Identifying healthcare as the most pertinent domain for their model, given the sector’s inefficiencies and the challenges in deriving insights from vast amounts of unstructured data, HARMAN faced the challenge of inadequate training data. However, they found a rich source of publicly reported data in clinical trial studies related to cancer, immune diseases, and heart diseases, utilizing this data to train HealthGPT.
Nevertheless, the dearth of quality data poses a significant obstacle. Models trained on skewed datasets can yield biased outcomes. “We have implemented mechanisms to detect and rectify biases in the data, not only during the data preparation phase but also during output scrutiny,” Dr. Ganesh added.
Rigorous testing frameworks, involving extensive querying, are indispensable to ensuring the integrity and privacy of processed data.
Responsible Hallucination Management
It is no secret that LLMs are susceptible to generating erroneous information, particularly in critical fields like life sciences, where inaccuracies can pose significant threats. To mitigate these risks, HARMAN employs a combination of automated mechanisms and human intervention.
HARMAN adopts a multifaceted approach to hallucination management. Initially, HealthGPT implements guardrails to regulate the extent of hallucination. “Initially, model accuracy stood at around 74%, but through continuous refinement, it has significantly improved, reaching over 85-90%,” Dr. Ganesh revealed.
Secondly, the model interface provides users with control over settings such as temperature and token number, enabling them to assess the degree of hallucinatory outputs. Higher temperatures amplify hallucination, while lower settings mitigate it. Thirdly, human oversight is incorporated, with medical professionals validating AI-generated results. This is complemented by a feedback mechanism for iterative refinement of the model. Additionally, the RAG feature adds references to responses for enhanced information credibility. Lastly, the system includes a benchmarking section that compares the model’s performance with other studies and models.
However, responsible AI lies at the core of HealthGPT’s success. “One of the reasons for opting for a private LLM is to ensure end-user control. Unlike models hosted on unfamiliar cloud instances, the HealthGPT model operates within the user’s own Virtual Private Cloud (VPC) and cloud environment,” Dr. Ganesh emphasized, highlighting how this approach enhances security and privacy by fine-tuning the model on the user’s data within their controlled environment.
Pre-fine-tuning checks are also implemented to detect anomalies, prioritizing privacy through automated mechanisms for handling Personally Identifiable Information (PII) and Protected Health Information (PHI).
HARMAN’s philosophy in the healthcare domain, as well as others, revolves around a human-centric approach. This entails comprehensively understanding problems and placing decision-makers at the forefront of solutions. This philosophy underpins HARMAN’s interactions with its diverse global customer base, ranging from companies in the proof-of-concept (POC) stage to those in more advanced stages of implementation.
What Lies Ahead
While not yet deployed for live customers, HealthGPT has demonstrated compelling proof-of-concept (POC) stories showcasing its potential applications across various domains. One POC illustrates HealthGPT’s utility for personalized data analysis across sectors, facilitating customization for individual needs such as drug discovery in pharmaceuticals.
Another story highlights the model’s effectiveness in extracting insights from medical instrument data, underscoring its capability to handle large-scale, structured information. A third user story features a pharmaceutical company leveraging the model to enhance drug discovery by integrating clinical trial data and information from sources like PubMed. IQVIA, Roche, Aetrex are among the prominent clients served by HARMAN.
“Currently, we are exploring Mistral AI’s Mixtral 7B for future iterations. Our goal is to continually push the boundaries of auto foundation models,” Dr. Ganesh stated.
HARMAN’s generative AI strategy entails integrating diverse data sources while prioritizing data quality and compliance with regulations like HIPAA. This necessitates extensive training for developers handling healthcare data. Concurrently, Dr. Ganesh and his team are striving to introduce multimodal features.
Regardless of the strength of the generative AI product, HARMAN’s strategy remains rooted in a human-centric philosophy. This approach involves understanding problems holistically and centering decision-makers in the solution process.
Furthermore, plans are underway to expand HealthGPT’s scope beyond healthcare into domains like manufacturing and IT management, aligning with HARMAN’s strategic objectives.
Conclusion:
Harman’s entry into the healthcare space with HealthGPT signifies a strategic move to address critical challenges in the industry, particularly concerning data privacy, quality, and reliability. By emphasizing a human-centric approach and leveraging innovative AI technologies, Harman is poised to disrupt not only the healthcare sector but also other industries, showcasing its commitment to driving impactful solutions in the market.
