In 2023, Tempus One set a standard for how providers and researchers interact with healthcare data, offering intuitive, AI-driven tools to query structured datasets and streamline workflows. Today, the platform takes a giant leap forward, integrating powerful generative AI capabilities that unlock insights from unstructured data, redefining what’s possible in clinical care and research.

The expanded Tempus One offering brings transformative new features that allow researchers to investigate unstructured data in our data analytics tool, Lens, as well as agents that provide Tempus team members the ability to query providers’ unstructured data records, enabling clinical trial eligibility and care gap assessment across sites. We are also introducing a newly enhanced Cohort Builder, designed to address specific challenges in healthcare and research workflows. These tools leverage Tempus’ cutting-edge GenAI platform and agentic workflows to deliver insights that were previously difficult—if not impossible—to access. Whether it’s helping providers explore care gaps for individual patients or enabling researchers to query millions of unstructured documents to uncover new novel drug development insights, Tempus One is empowering users like never before.

In this blog post, we’ll dive into these new capabilities, including how researchers can use Data Exploration in Lens to interrogate datasets at scale, how recent enhancements to Cohort Builder are accelerating real-world evidence generation, how Patient Query helps providers optimize care pathways, and how Patient Timeline can provide a comprehensive, longitudinal view of a patient’s full medical history. With these tools, Tempus is transforming the role of AI in healthcare, making precision medicine and breakthrough discoveries more accessible for everyone.

We’ve developed these agents under a governance program that is intentional about privacy considerations and the responsible development and use of AI in healthcare. We seek to develop safe, effective, and transparent AI, especially when its outputs are relied on in a way that will impact patient care. And, of course, we follow FDA’s rules for medical devices, if the agent is a device.

Data Exploration

Tempus has amassed one of the largest multimodal datasets and built Lens so that researchers can make sense of that data. Our customers who license our de-identified datasets have access to structured data, which is a variety of data elements that are meticulously curated into tables by our abstraction team to enable easy analysis. This process creates a detailed, structured representation of patient information using medically validated rules and guidelines. It also ensures that quality real-world data (RWD) is structured and available to our customers.

But our dataset goes much further beyond the typical categories of data found in de-identified, structured records. Users may want to analyze a novel data element within the unstructured clinical documents. Tempus One’s newest functionality in Lens allows researchers to query the following items from a de-identified dataset that are not usually available in structured research records:

• Adverse events and reported symptoms, including a privacy-enabling time period reflecting when they occurred
• Imaging tests performed, including the protocols and contrast agents used, and the time period of the test
• Radiological report information, including the lesion size, reports with no findings, and incidental findings
• Comorbidities
• Innovative/novel testing (e.g. liquid biopsy test)
• Traditional diagnostics (e.g. antigen testing)
• Off-label treatments

Now, users can use Tempus One to do “cohort exploration,” specifically asking questions iteratively and reviewing aggregate cohort characteristics from Tempus’ structured data and unstructured clinical documents. This new capability enables a new form of research that was hindered by the cost and time delays of data abstraction.

Cohort Builder

In early 2024, we announced that we had integrated Tempus One in Lens, unveiling a whole new, AI-enabled functionality for our research collaborators. With that launch, we introduced Cohort Builder, which helps define patient cohorts of interest by leveraging the platform’s rich molecular and clinical filters, including, cancer diagnosis, medications, genomic alterations, staging, and more.

Since then, we’ve continually invested in enhancing the no code interface in which users can narrow down to specific cohorts that will answer their research questions. Anyone that has worked with RWD knows it requires a certain level of training and familiarization with the many acronyms, short hand terminology, and nomenclature that you encounter across different databases. To give you an idea, below is just one example of what one query can look like:

At Tempus, we understand that our data is valuable only if you are able to easily read and analyze it. Cohort Builder creates queries against Tempus’ data using each user’s language and terminology. It is designed to meet users where they are regardless of their background experience and familiarity with Lens, Tempus data, and RWD overall.

How do we do this? We integrate natural language interface into Lens, allowing users to use a hybrid of text queries and drag/drop/edit to query and analyze our RWD database – something that is only possible with the advent of LLMs.

We found the following as part of early testing:

• In the past 90 days, Lens users who engaged with this feature generated over 400 queries with GenAI. This is a powerful feature for new users who may not be conversant with the complex data model.
• Beta testing results indicated that users generally found the tool useful. Users evaluated responses for a total of about two thousand queries and rated the feature highly. We have found that an expert Lens user can create a filter using Cohort Builder in half the time it takes to develop the same query manually.

The next step is to integrate advanced business intelligence capabilities so that users can generate deeper insights from these cohorts.

Patient Query

In 2019, Tempus launched its TIME Trial clinical trial matching program, which works to bring clinical trials to patients in communities across the U.S. Using a standardized operational framework, TIME seamlessly connects patients and their providers with cutting-edge clinical trials so that patients can participate in trials closer to home. TIME also assists health systems in identifying patients eligible for clinical trials through enhanced patient matching services, alleviating some of the significant burden associated with pre-screening.

Patient Query enables our TIME team to rapidly query unstructured patient data. This helps to unlock more detailed patient information that is not typically captured in structured data but is critical to assessing trial eligibility, such as hospice status, presence of exclusionary comorbidities, current line of therapy, and more. This data provides insights into the patient population at providers across the TIME Network, informs our work with trial sponsors regarding enrollment potential, as well match specific patients with clinical trials with higher fidelity and quality.

Our new ability to apply LLMs to evaluate hundreds or thousands of patients against a specific clinical trial’s eligibility criteria significantly cuts down on the time our registered nurses need to spend to manually screen that full list of patients and instead prioritizes those patients that are most likely to be a match for that study.

How it works:

Historically, Tempus has used a two step approach to patient matching for clinical trials.

1. First primarily structured data is used to identify patients that may meet high level trial eligibility (i.e. tumor type, stage, biomarker).
2. Then, TIME oncology nurses utilize all available patient data to screen against full trial eligibility criteria. For patients at a provider site with an EMR integration, this can include a substantial amount of unstructured data (i.e. progress notes) that must be manually reviewed to ultimately determine if a patient is a match for a trial.

With the implementation of Patient Query in TIME, the list of patients meeting the high-level trial criteria generated in step 1 is then run against LLM queries assessing additional eligibility requirements. Each patient is given a “match score” based on the number of queries satisfied (0-1, 1 = satisfied all queries). These match scores are returned to the queue, allowing the registered nurses to then sort the patients based on those that satisfy the most queries (i.e. those predicted to be most likely matches).

In one trial where Patient Query is currently implemented, we found that the patient match rate was 66% among patients with a match score of 1. Based on this, we implemented a screening threshold where only patients with a match score of 0.8 or higher would be reviewed by a nurse. Since then, we have already saved over 45 hours of nurse screening time for this single trial.

The incorporation of LLMs into our standard patient screening process facilitates faster provider notification of high-quality patient matches for clinical trials and dramatically decreases the human effort required. This speed improves the likelihood that a patient can actually enroll on a clinical trial and hasn’t already started a new line of therapy instead, and the increased efficiency allows the same number of nurses to review patients across a larger portfolio of clinical trials.

In addition to assessing which patients may be a match for a specific clinical trial, this agent can also assess which patients might be experiencing a care gap. Often, finding gaps in care for patients requires assessing data that is both reliably captured in the EMR as structured data and interpreting unstructured documents to determine if guidelines are met. This tool now allows our team to query a provider site’s unstructured data to further qualify those patients for a care gap.

Patient Timeline

For many cancer patients, their care journey collects hundreds of documents across providers and specialists, making it incredibly difficult for a treating physician to have one holistic view of their patient. Patient Timeline aims to provide a comprehensive, longitudinal view of a patient’s medical history by integrating all of the patient’s data into one place. This timeline serves as a dynamic, evolving tool to track key clinical events over time, supporting healthcare providers in delivering better personalized care.

Using an LLM-based model, we synthesize a wide range of data sources including patient progress notes and other documents into a cohesive, easy-to-navigate timeline. To develop a patient timeline, we leverage long-context LLMs, Gemini-1.5-pro, to accurately identify key clinical events and important details, like event dates and relationships between events such as initial diagnosis, symptom onset, treatment at all stages, medication and side effects, and comorbidity trends. We enrich each event with relevant context extracted via the LLM from unstructured text.

The timeline is generated in a json format with loose contract in a semistructured format that allows further advanced features. For example, allowing the clinician to filter by event type (e.g., medications), or by time window (e.g., last 6 months, past year). An incremental approach to generating the Patient Timeline ensures that updates to a patient’s clinical history are efficiently incorporated into the timeline as new data becomes available.

In addition to the longitudinal view, a separate agent queries this data to answer explicit questions posed by a physician either within individual events or across the entire timeline for specific terms or conditions, leveraging LLM-powered capabilities. All of this functionality is made available through the Tempus One interface in Hub.

To understand the performance of the generated patient timeline, we have compared the generated timeline against patient abstraction data. We have evaluated how well the system integrates unstructured data and ensures the completeness of the patient timeline, including checking for missing or inaccurate data. We also assess the accuracy of the LLM in extracting key events from the unstructured notes evaluating for facts and contradictions. In addition, we also evaluate the generated timelines for consistency across multiple runs.

Tempus Prior Authorization

Physicians and care teams spend countless hours drafting the prior authorization letters that must be submitted to insurance companies to request coverage for treatments for their patients. To alleviate this large administrative burden, Tempus developed a tool that drafts those prior authorization letters automatically in Hub. The new tool generates a request letter using data about a patient, drug label, and payer policy, so that physicians can waste less time on administrative tasks and more patients get access to clinically necessary medications.

Conclusion

We’ve only just begun to unlock the promise that LLMs hold in better supporting our physician and research partners, as well as our own team. These are just a few of the newest tools, and we look forward to building on this work to further hone them while also creating new solutions that will make personalized care a reality for so many more patients. Stay tuned.