Triomics, an oncology-focused artificial intelligence company, raised $15 million to help automate cancer providers’ workflows.
Investors supporting the round include Lightspeed, Nexus Venture Partners, General Catalyst and Y Combinator.
Triomics co-founders Sarim Khan and Hrituraj Singh were college friends who later worked as an MIT biotech researcher and Adobe AI researcher, respectively. While software exists to quickly analyze the 20% of medical data that are stored in a uniform, structured manner, like a patient’s demographics or laboratory values, they realized recent advances in generative AI created the possibility of similarly analyzing the 80% of medical data that exist in an unstructured format, like a doctor’s free-text note.
Triomics’ preeminent technology is a large language model called OncoLLM. The model can mine unstructured data sitting in patients’ medical records using institution-specific inputs and specific use cases to perform a variety of tasks for providers.
Triomics touts the model’s ability to match patients with clinical trials, which the National Cancer Institute says can be used to manage side effects of cancer treatment and test new treatments, among other aims. An April study published in the Journal of Clinical Oncology says that only about 20% of cancer patients are enrolled in a clinical trial.
In one example, the model, within minutes, matched 90% of patients with a clinical trial by searching patient records, which would have taken healthcare staff hours of manual processing per patient due in part to strict inclusion and exclusion criteria for trials, according to the company.
It also extracted structured data points from unstructured notes at similar or higher accuracy to proprietary models like GPT4 or Claude while being 40 times cheaper, Triomics' executives said.
What sets Triomics apart, the company says, is its extensive academic partnerships and published literature. One study done in collaboration with the Medical College of Wisconsin on clinical trial matching, is said to be the first end-to-end large-scale empirical evaluation of clinical trial matching using real world electronic health records.
The study found that OncoLLM exhibited a novel trial ranking method whereby it provides a ranked list of trials for which patients would be best matched. The model can be used in the opposite direction to identify patient cohorts eligible for clinical trials. The researchers found that OncoLLM outperforms GPT-3.5 but also matches the performance of qualified medical doctors.
Triomics says OncoLLM performs almost as well as Azure OpenAI’s GPT-4.
Since the study’s publishing, Triomics has active pilots in several cancer centers with data forthcoming by the end of the year, a Triomics spokesperson told Fierce Healthcare.
The study also says that running GPT-4 costs about $6,055 per hour, while OncoLLM costs about $170 per hour. The company says the significant price differential may be beneficial to some systems. It also touts the privacy benefits of using a smaller model like OncoLLM compared with proprietary models like GPT-4.