How HARMAN is Solving Healthcare Problems with Generative AI
In the bustling arena of technological innovations, HARMAN, a Samsung subsidiary, is carving a niche with its groundbreaking Generative AI model, HealthGPT, tailored for the healthcare sector. HealthGPT, built on the open-source Falcon model from TII and leveraging Llama 2 for its latest version, HealthGPT Chat, symbolizes a forward leap in utilizing AI for healthcare applications. The complexity and sensitivity of data in healthcare present unique challenges, prompting HARMAN to introduce a private Large Language Model (LLM) to address privacy and security concerns rampant in the industry.
Dr. Jai Ganesh, chief product officer at HARMAN, shared insights with AIM, revealing the impetus for developing HealthGPT. “The decision to create a private LLM sprung from apprehensions regarding data privacy and structural limitations inherent in public models. As enterprises hesitated to transfer sensitive data to third parties, it became essential to devise a solution that alleviated these concerns without compromising on functionality,” explains Dr. Ganesh.
HARMAN’s strategic pivot to private models coincided with the release of several open-source foundational models in early 2023, a period marked by a significant movement towards accessibility in AI technologies. However, instead of constructing a foundational model from scratch, HARMAN opted to harness the potential of existing open-source models—a decision influenced by the prohibitive costs and resources required for such an endeavor.
“Our focus gravitated towards healthcare due to the domain’s urgent need for efficient data analysis and decision-making support,” Ganesh added. Utilizing publicly available clinical trial data, HealthGPT was trained to tackle issues in cancer research, immune diseases, and heart conditions, among others. Rigorous testing and bias correction mechanisms ensure the model’s outputs are reliable and unbiased, a critical factor given the life-or-death stakes in healthcare applications.
HARMAN employs a multifaceted approach to manage the risks of inaccuracies and hallucinations in AI-generated outputs. Initially achieving an accuracy rate of approximately 74%, the models have since been refined to achieve a remarkable accuracy rate exceeding 85-90%. The model incorporates guardrails against hallucinations, user controls for temperature and token settings, and human-in-the-loop validation to ensure reliability and relevance of its outputs.
Central to HealthGPT’s success is HARMAN’s commitment to responsible AI. HealthGPT operates within the user’s Virtual Private Cloud (VPC), ensuring data privacy and security by fine-tuning models on user-controlled environments. Pre-fine-tuning checks for anomalies and automated handling of Personally Identifiable Information (PII) and Protected Health Information (PHI) bolster the model’s privacy safeguards.
HealthGPT’s proof-of-concept (POC) stories highlight its versatile applications, from personalized data analysis across sectors, including pharmaceuticals and medical device data interpretation, to bolstering drug discovery efforts through the integration of clinical trial data and authoritative sources like PubMed.
Looking ahead, HARMAN’s exploration of Mistral AI’s Mixtral 7B for future iterations of HealthGPT represents a commitment to enhancing the model’s capabilities. The company aims to integrate more diverse data sources, ensure data compliance, and introduce multimodal features, maintaining a staunch adherence to human-centric philosophy in AI development.
While initially focused on healthcare, plans are underway to expand HealthGPT’s applications to manufacturing and IT management, showcasing HARMAN’s broader strategic ambitions. In a rapidly evolving technological landscape, HARMAN’s journey with HealthGPT underscores a significant contribution towards solving healthcare problems through the innovative use of generative AI.
