Catching Cancer Cachexia Earlier: A Multimodal AI Biomarker Built from Routine Clinical Data
A validation study showing that combining clinical notes, lab tests, imaging, and routine patient data produces consistently better predictions of cancer cachexia than any single data modality, across multiple cancer types.
The problem: a syndrome that hides in plain sight
Most people have never heard of cancer cachexia, but it shapes the experience of millions of cancer patients. It's a wasting syndrome marked by progressive loss of muscle, weight, and strength that develops alongside many cancers. It makes chemotherapy harder to tolerate, surgery riskier, daily life more exhausting, and survival shorter.
Cachexia is common. It affects up to 70% of patients with pancreatic and gastroesophageal cancers, and roughly half of those with blood, colorectal, and lung cancers. Yet despite how common and serious it is, there is no single test that reliably catches it early.
The reasons are practical. Blood markers like albumin and C-reactive protein, weight measurements, and muscle area measured on CT scans each tells part of the story, but no single signal is enough on its own. Existing scoring systems try to combine a few of these signals using fixed numerical cutoffs, but those cutoffs don't transfer well across different cancer or patient populations. The more thorough assessment tools take too long to be incorporated in clinical settings.
The result: cachexia is usually diagnosed after it has become visibly obvious, long past the window when nutritional support, exercise, and other interventions could have helped most.
Why a multimodal approach
The core insight of this work is that cachexia is a multidimensional problem and deserves a multidimensional answer. It doesn't show up in just one place in the medical record. It leaves traces in the lab work, in the imaging, in what a patient tells and in what their oncologist writes after a clinic visit, and in the slow shifts in their weight over time.
What's needed is a support system that reads everything at once and learns the patterns from the combination, the way an experienced clinician would.
What we built
We built an AI system that pulls together four different kinds of information from a patient's medical record at the time of cancer diagnosis:
- Basic clinical data: age, height, weight, BMI, and cancer stage.
- Lab results: standard blood tests that reflect nutrition and inflammation.
- CT imaging: assessment of how much muscle the patient has.
- Doctors' notes: the free-text observations written during clinic visits, which often describe symptoms (appetite loss, fatigue, weight history) that don't show up in numbers.
To make the notes useful to the model, we used large language models to read each patient's notes and pull out structured information about cachexia-related symptoms. The framework we built then learns how to combine all four data types or data modalities into a single prediction for each patient.
Validating the approach across cancer types
For this validation study, we tested the system on three patient cohorts:
- 236 pancreatic cancer patients from the Florida Pancreas Collaborative: a multi-hospital network spanning several institutions across the state of Florida.
- 131 pancreatic cancer patients treated at Moffitt Cancer Center, where the full set of modalities (including clinical notes) were available.
- 135 ovarian cancer patients: a different type entirely, used to test whether the pattern generalizes beyond pancreatic cancer.
The same model architecture and approach were applied to all three cohorts. The question was “does the use of multimodal data help with improved detection of cancer cachexia?”
What the results show
The central finding of this validation is that adding data modalities consistently improved predictions across every cohort and every metric we measured.
Here's how accuracy progressed as we added each type of information:
| Data modalities | Pancreatic (PDAC) n = 236 | Pancreatic (PDAC) Moffitt n = 131 | Ovarian n = 135 |
|---|---|---|---|
| Clinical data only | 59.5% | 59.6% | 53.2% |
| + Muscle measurements from CT | 60.5% | 59.0% | 56.7% |
| + Lab results | 65.0% | 66.5% | 57.5% |
| + Symptom information from clinical notes | 67.4% | 71.8% | not available |
The same upward trend appears in two other measures of model quality. The AUC, a standard score for how well a model separates patients with and without cancer cachexia, climbed from 62% (clinical only) to 72% (full multimodal) in the larger pancreatic cohort, and from 63% to 78% in the Moffitt subset. The F1 score showed the same pattern, rising from 67% to 75% in the larger cohort and from 65% to 77% in the Moffitt one. The ovarian cancer cohort showed a similar trend with AUC improving from 52% (clinical only) to 60% (clinical + muscle measurements + lab) and F1 improving from 58% to 64% with the combination of three data modalities excluding notes.
Several findings stood out:
- The pattern was consistent across cancer types. Pancreatic and ovarian cancers are biologically and clinically different, but the multimodal benefit appeared in both. This is the strongest argument that the multimodal approach is capturing information about cachexia from all types of data modalities included.
- Lab data and clinical notes contributed the most. Adding lab results produced a sizable jump in every cohort. Adding the symptom information extracted from clinical notes, where it was available, produced another. Muscle measurements added considerable gains in the ovarian cancer cohort but smaller gains in the pancreatic (PDAC) cohort, highlighting how different types of cancers draw meaningful information from the various data modalities combined.
- It gives a personalized answer. Instead of applying the same numerical cutoff to every patient, our framework produces a prediction tailored to the individual, accounting for their age, ethnicity, cancer type, stage, and other characteristics.
- It uses data that's already being collected. Nothing in the pipeline requires special collection protocols, new tests, or extra patient visits. The system draws on what's already in the electronic health record from standard cancer care.
Why this matters
By the time the syndrome is recognized, the best opportunities for intervention have passed. A tool that can flag at-risk patients on the day of cancer diagnosis by reading across all the information the hospital is already gathering opens up a different timeline. Nutritional counseling, physical therapy, appetite-supporting medications, and closer monitoring can all begin earlier, when they're more likely to help.
The broader lesson goes beyond cachexia. Multimodal AI is what cancer medicine has always actually needed, because cancer has never been a single-data-source problem. The information that defines a patient's situation is scattered across imaging, labs, pathology, genomics, and the words their care team writes about them. Tools that bring all of that together will look more like the way clinicians already think and that's the direction this work is pushing.
What this validation adds to the conversation is concrete evidence that the multimodal benefit is real, measurable, and consistent across cancer types. It's not a coincidence of a single dataset or a single disease. The same pattern — each modality adding something — appears in pancreatic cancer across multiple hospitals, in pancreatic patients at a single center with the richest data, and in ovarian cancer as a different disease entirely.
What's next
This study validates multimodal learning in two cancer types and that direction needs to keep going across more cancer types, more institutions, and over time. We're working toward longitudinal modeling using repeated measurements to track how a patient's risk evolves rather than capturing a single snapshot at diagnosis. Ultimately, the goal is a tool that supports oncologists in everyday practice, helping them catch cachexia early enough to actually do something about it.