Charting the patient journey: Unlocking new insights with longitudinal multimodal data

Melisa Tucker: A primary challenge in leveraging real-world data is that patient information is often fragmented across different health systems and difficult to link. Tempus overcomes this by building a continuous record for each patient, anchored by a master patient identifier that tracks individuals as they move between sites of care.

We center our platform on direct data connections at the source, which minimizes the patient drop-off that often occurs when linking separate datasets. This preserves the integrity and scale of the cohort. We integrate several key data types, such as:

• Molecular data: Generated in-house from our comprehensive portfolio of sequencing tests, providing multiple snapshots of a patient’s molecular profile as their disease and treatment evolve.
• Clinical data: Sourced from direct electronic health record (EHR) integrations with thousands of care sites, capturing structured data, unstructured notes, and scanned documents on an ongoing basis.
• Imaging data: Including a substantial library of H&E slides and a growing collection of DICOM radiology images, linked at the patient level to their clinical and molecular information.

This raw, multimodal data undergoes extensive harmonization and curation to create analysis-ready variables, including disease characteristics, biomarker status, lines of therapy, and key outcomes like progression, treatment response, and overall survival. The result is a rich, longitudinal view of each patient’s journey.

Tim Hagerty: This is where the data becomes incredibly powerful. We know that cancer is not a static target; it’s an adaptive, plastic system that changes in response to different drugs and stimuli. Understanding these changes is critical for developing more effective therapies.

We use this longitudinal data to investigate these dynamics through analyses we conduct internally at Tempus. For example:

• Mechanisms of resistance in triple-negative breast cancer (TNBC): By comparing biopsies taken pre-treatment and at progression, our analysis revealed that different mutations were associated with primary versus acquired resistance. PIK3CA mutations were more common in patients with primary resistance (progression within three months), while PTEN and KMT2C mutations emerged more frequently in those with acquired resistance.¹
• Tumor microenvironment (TME) remodeling: In metastatic TNBC, transcriptomic analysis of post-progression samples suggested that resistance to sacituzumab govitecan could be driven by remodeling within the TME, highlighting our ability to analyze the entire tumor ecosystem.¹
• Prognosis in non-small cell lung cancer (NSCLC): An analysis of patients with EGFR-positive NSCLC showed that acquired RB1 loss-of-function mutations, which increased in prevalence from 6.8% pre-treatment to 11.7% post-osimertinib, were associated with significantly worse real-world overall survival. This demonstrates our ability to not only identify molecular shifts but also quantify their clinical impact.¹
• Treatment paradigm evolution: Our platform can track the real-world adoption of new therapies over time. Following the approval of nivolumab plus chemotherapy in frontline esophageal cancer, our data showed a clear shift in treatment patterns and demonstrated similar uptake rates across academic and community sites.¹

Melisa Tucker: The same rich, multimodal data that powers these retrospective analyses also forms the foundation for our work in predictive modeling. I believe the primary challenge in building effective AI models in oncology is not a lack of data volume, but a lack of deeply integrated, high-quality data that captures the complete patient picture. The hard part is being able to integrate that patient’s full picture and have all the relevant outcomes, clinical characteristics, and features attached at the patient level.

Our goal with our foundation model work is to achieve an understanding of disease that isn’t possible from a single data modality alone. By integrating clinical, molecular, and imaging data over time, we aim to build models that can predict what might happen next in a patient’s journey. The increasing use of ctDNA for minimal residual disease monitoring is transforming this capability, moving from discrete snapshots to a more continuous signal of a patient’s molecular status. This allows for real-time monitoring and creates the potential to predict disease changes before they are clinically apparent.