This article has been updated to clarify how Layer Health is working with health systems.
NEW YORK – Layer Health is developing its large language model (LLM)-based data abstraction platform as a tool to help doctors make individualized treatment decisions for patients using the totality of their medical history, including data from unstructured notes.
The Boston-based technology company inked an agreement last week with the American Cancer Society to apply its LLM platform to abstract data from thousands of medical charts of patients enrolled in ACS studies and glean new insights that can improve drug discovery and cancer treatment guidelines. The firm's ambitions also include bringing this technology into the exam room to help doctors personalize patient care by harnessing test results, biomarkers, disease symptoms, and other clues hidden in electronic health records (EHRs).
Layer CEO David Sontag, a professor of computer science at Massachusetts Institute of Technology, worked on artificial intelligence in healthcare for about 15 years before founding the company. He said one of the biggest challenges he and his colleagues encountered when developing algorithms to identify the best medicines for patients was accessing unstructured notes in EHRs. "That information problem is critical because the data you need … is typically found in clinical notes that doctors write, not in any structured data that would be available in EHRs," Sontag said.
While Sontag had long been thinking about this problem, the impetus to start up a company dedicated to developing an AI solution came during a personal encounter with cancer. Sontag's mother had been diagnosed with smoldering multiple myeloma, a blood and bone marrow disorder that can advance to multiple myeloma, and she was prescribed the standard treatment: watchful waiting. Patients with smoldering myeloma have a 10 percent annual risk of progressing to multiple myeloma. If they don't progress after five years, the risk drops to 3 percent per year for another five years, and then it is 1 percent per year after that. Because there are no symptoms and a significant proportion of patients will never develop the more aggressive form of the disease, doctors prefer to monitor patients for progression rather than intervening early.
"My mom got her lab panel, a bone marrow biopsy … and that, together with the blood biomarkers, led the clinical team to [recommend] not to start treatment yet," Sontag said. However, after three years, she developed heart failure due to cardiac amyloidosis, a complication of multiple myeloma caused by a buildup of abnormal immunoglobulin light chain proteins in the heart. Doctors immediately began treating Sontag's mother for multiple myeloma, but it was too late, and she died a month later.
"It all became so clear to me in that moment," Sontag said. "There was an existing staging algorithm that looked at a single number from her bone marrow biopsy and said the right treatment for her was the same as the right treatment for everyone else, which was [watchful waiting]. This is the opposite of precision medicine."
Sontag reflected that with the inclusion of genomic data, a more accurate prediction of his mother's course of disease could have been made. "I've had students who work with single-cell RNA-seq data from patients with multiple myeloma, and with that data, you can really distinguish who's going to progress and who is not going to progress."
In addition to genomics, another type of data that could have yielded clues about Sontag's mother's progression to cancer would have been the symptoms she experienced before being diagnosed with cardiac amyloidosis, such as bruising in her eye and arthritis in her hand. These symptoms, in retrospect, were signs of the disease, the amyloidosis specialist said.
"The only way to tackle this problem is to look holistically at the patient, so we can connect the dots across, for example, what different providers are saying about the patient and the symptoms and piece together a story for the oncologist that helps them see the bigger picture," Sontag said. "At the end of the day, having the right information about the patient can actually guide the right treatment decision, but today's electronic medical records aren't built to be able to do that in any personalized way."
Sontag said he and Layer Cofounder Steven Horng, an emergency physician at Beth Israel Deaconess Medical Center in Boston and a clinical informatician at Harvard University, worked with their students for more than a decade to reinvent the functionality of EHRs. They developed machine-learning algorithms that could continuously predict in real time the key variables in a patient's condition and then added a clinical decision support tool that answers clinical questions using an evidence-based approach on top of that machine-learning layer.
"You could imagine having a clinical pathway that's relevant for this patient, surfaced for the clinician at the right time based on information our algorithms are outputting," Sontag said. That vision, of having a "layer" of AI-surfaced information available to physicians for decision support, inspired the company's mission and name.
In a pilot study involving breast cancer patients, Layer's data abstraction platform was able to extract complex information from patients' charts with 95 percent to 100 percent accuracy. This study forms the basis of the company's collaboration with ACS, in which, over the course of several years, it hopes to analyze patients' charts, including those enrolled in ACS's Cancer Prevention Study-3, and discover what the optimal first-line therapy is for each patient. Those insights "will then lead to new treatment guidelines that are going to directly affect patient care," Sontag hopes.
Sontag added that within the collaboration, the LLM will be constrained to provide only direct quotes from patients' records as responses, not summaries or paraphrases. That innovation, he said, means that "it's no longer possible for the model to just hallucinate a quote because we can actually explicitly enforce that it has to be character by character from the record."
In addition to ongoing development of the platform as a clinical decision support tool and its work with the ACS, Layer is also hoping to demonstrate the technology's performance in specific use cases including matching patients to clinical trials and discovering biomarkers for drug development. The company is also working with health systems to evaluate the platform's ability to automate submissions to clinical registries and, longer term, make sure cancer patients are receiving appropriate treatments.