Our Model
OncoLLM
TM
Generative artificial intelligence (gen AI), brought mainstream by OpenAI’s ChatGPT, is a paradigm-shifting technology. But using this technology to improve cancer care is not as simple as buying a $20 monthly subscription. Whether hidden in free-text notes or stored in complicated oncologies, a large amount of clinical and genomic oncology data remains difficult to use for research and care.
Oncology Performance:
State-of-the art, generalized LLM vs OncoLLM
.jpg)
As off-the-shelf gen AI solutions need to “know” a little bit about everything, from Shakespeare’s 800,000 words to the 17,000 species of butterflies, their training datasets are necessarily humongous, driving up costs.
This breadth of knowledge means that depth of knowledge must be sacrificed, an obvious problem in oncology, where each and every detail could be life-altering. Inherently, the size of generalized large language models (LLMs) requires offsite storage, hosting and data transfer, creating major security and compliance risks.
Our team of over 50 technologists and clinicians, including experts in artificial intelligence (AI), clinical workflows, and oncology informatics, helps cancer centers leverage their own data to train a bespoke model. This model, which can be hosted locally, is uniquely able to streamline pain points at each institution.
OUR SOFTWARE
However, and in contrast to other gen AI providers, we don’t simply hand off a model and leave.
Rather, we’ve designed intuitive, workflow-specific software that integrates seamlessly with EHRs using standards like FHIR, leveraging the model’s input to sift through patient data and execute desired workflows. Of note, this process flow mimics, rather than alters, existing clinical workflows.
OUR SOFTWARE
However, and in contrast to other gen AI providers, we don’t simply hand off a model and leave.
Rather, we’ve designed intuitive, workflow-specific software that integrates seamlessly with EHRs using standards like FHIR, leveraging the model’s input to sift through patient data and execute desired workflows. Of note, this process flow mimics, rather than alters, existing clinical workflows.
.png)
.png)
OUR RESULTS
OncoLLM, developed in collaboration with a cancer center, was able to outperform larger open-source and proprietary LLMs (link) at patient-trial matching. Its performance even rivaled qualified medical experts and GPT-4, despite being much smaller and 35 times less expensive. Since publication, we have developed another variant of OncoLLM (70B) with accuracy surpassing qualified medical experts and GPT-4.
OUR SOFTWARE
However, and in contrast to other gen AI providers, we don’t simply hand off a model and leave.
Rather, we’ve designed intuitive, workflow-specific software that connects seamlessly with EHRs using standards like FHIR, leveraging the model’s input to sift through patient data and execute desired workflows. Of note, this process flow integrates with, rather than alters, existing clinical workflows.
Plus, our software shows its work to protect against "hallucination," a known challenge for generative AI solutions. For example, our software will cite the patient chart to show how a patient meets or fails each inclusion and exclusion criterion.
Application layer
Model + Software = Value
Integrated workflow softwares that enables the end user to carry out a task such as patient pre-screening, data curation at the scale.
End User
(Model + Software + Service = Adoption)
Stakeholders, including trial coordinators, registered nurses and registrars, engage with the application layer to ajudicate the model’s output prior to using the data
OncoLLM
Model only = Potential
An oncology aligned model that run inside the firewall, processing free-text data with high accuracy at a fraction of the cost.
CONTACT US
888-ONCO-LLM
(888-662-6556)
601 Montgomery St, Suite 1100
San Francisco, CA 94111