Skip to Main Content

An official website of the United States government

Principal Investigator
Ananya Singh
Indian Institute of Technology
Position Title
About this CDAS Project
PLCO (Learn more about this study)
Project ID
Initial CDAS Request Approval
Jun 7, 2024
Develop an end-to-end LLM-powered application tailored for precise care of cancer patients with advance features
Our model includes features like:
Text Generation: When presented with structured and unstructured data from oncology patient records, the LLM can generate detailed descriptions of tumor characteristics, treatment modalities, and prognosis. This information can aid clinicians in understanding the patient's condition and making informed decisions regarding their care.

Summarization: Given a lengthy oncology research paper or a patient's extensive medical history, the LLM can generate succinct summaries highlighting key findings, treatment recommendations, and significant events. These summaries provide clinicians with a high-level overview of relevant information, saving time and improving efficiency in decision-making.

Question Answering (Q&A): When presented with questions such as "What are the common side effects of chemotherapy in breast cancer patients?" or "What is the recommended treatment for stage III lung cancer?", the LLM can provide accurate and evidence-based answers sourced from oncology literature, clinical guidelines, and real-world patient data.

The implementation of our solution revolves around five main components
Healthcare Cancer patient dataset: We will utilize Healthcare dataset containing patient information, treatment histories, family history and clinical outcomes related to cancer care. This dataset will serve as the foundation for training and testing our LLM-powered application.

LLM (Llama / GPT) model: We will employ a Llama-based fine-tuned model as the core engine of our application. This model will be optimized for cancer care applications and equipped with features for text generation, summarization, and Q&A.
Alignment or Fine-tuning of Model Methods: Our approach involves meticulous alignment of the model on the Oncology dataset to meet clinical requirements and continuously refine and optimize the model's performance.This includes
Prompt engineering techniques such as zero-shot, one-shot, few-shot inference, Generative configuration (Temprature, Top P and K Sampling, Max new tokens).
Fine-tuning methodologies: Parameter Efficient fine-tuning using LoRA (Low Rank Adaption).
Clinical feedback mechanisms using Reinforcement Learning with Proximal Policy Optimization (PPO).

The applications of our LLM-powered application are vast and impactful:

Automated Prompt Generation: Our application automates the generation of LLM prompts from the Cancer patient dataset, enabling clinicians to access relevant information quickly and accurately.

Real-time Decision Support: By presenting generated prompts alongside patient data, our application provides clinicians with real-time insights and decision support capabilities, facilitating more informed and personalized care for cancer patients.

Enhanced Efficiency: With features such as text generation, summarization, and Q&A, our application streamlines clinical workflows and reduces the manual burden on healthcare providers, leading to increased efficiency and productivity.

Auto-generation of LLM prompts from the specifications of the oncology data model aiming to streamline the process of data extraction
We can integrate AI and LLM with our data to achieve:
1.Precision care with the fusion of AI and LLM
2.Extracting precise and granular information through open-source LLM model like Llama
3.Extract insights from longitudinal journey of cancer patient and delving in their family health histories
4.Revolutionizing cancer care by offering tailored interventions and improving health outcomes.
Our model includes features like:
Text Generation
Question Answering (Q&A)


GE Healthcare