# Multimodal Federated Learning for Pan-Cancer Detection of Cachexia and Early Detection of Pancreatic Cancer

> **NIH NIH P30** · H. LEE MOFFITT CANCER CTR & RES INST · 2024 · $202,200

## Abstract

Cancer-associated cachexia and pancreatic ductal adenocarcinoma (PDAC) represent two significant
challenges in oncology. Cachexia, a severe syndrome characterized by weight loss and muscle atrophy,
dramatically worsens patient outcomes across various cancers. Early detection of cachexia is essential for
improving quality of life and survival rates. Concurrently, PDAC is one of the deadliest forms of cancer, often
arising from intraductal papillary mucinous neoplasms (IPMNs), which are pancreatic cysts with malignant
potential. Accurate risk stratification of IPMNs is crucial for early intervention. Both conditions necessitate
advanced predictive models that leverage diverse data sources while ensuring patient privacy.
 The overarching goal of this proposal is to develop and validate privacy-preserving multimodal machine/deep
learning (ML/DL) models for the early detection of cachexia across multiple cancer types and risk stratification
of PDAC from IPMNs using clinical, imaging, genetic mutational, and social determinants of health (SDOH) data.
To access large-scale representative data while respecting the privacy and the security of sensitive patient data,
this proposal leverages Federated Learning (FL). This advanced ML paradigm enables the development of
robust predictive models across multiple institutions without centralizing data. FL allows each participating
institution to train models locally on their own datasets, sharing only model parameters (e.g., weights and
gradients) with a central server. This approach maintains data confidentiality and complies with privacy
regulations such as HIPAA and GDPR. By aggregating model updates from multiple sites, FL enhances the
diversity and generalizability of the models, reducing biases and improving performance across different
populations.
 Aim 1: Develop a multimodal artificial intelligence (AI) model for the early detection of cachexia in cancer
patients using clinical data, imaging biomarkers from abdominal computed tomography (CT) scans, and SDOH
for pancreatic, colorectal, gastrointestinal, and ovarian cancers.
 Aim 2: Develop a multimodal predictive model for risk stratification of IPMNs to identify those most likely to
progress to invasive pancreatic malignancy using imaging, clinical, patient-specific risk factors, SDOH,
histopathology, and genetic mutational data.
 The federated model’s performance will be benchmarked between sites and the centralized model using F1
score, AUROC, and accuracy metrics. The successful completion of these aims will enhance the early detection
of cachexia and risk stratification for PDAC, ultimately leading to better patient outcomes. We are fully prepared
to comply with all the administrative requirements outlined in the FOA. Our team has deployed NVIDIA-FLARE
and OpenFL with PyTorch and NumPy on Moffitt’s AI cluster and is conducting FL studies. Our NVIDIA-FLARE
environment will support FL model training as client and server.

## Key facts

- **NIH application ID:** 11159988
- **Project number:** 3P30CA076292-26S3
- **Recipient organization:** H. LEE MOFFITT CANCER CTR & RES INST
- **Principal Investigator:** John L. Cleveland
- **Activity code:** P30 (R01, R21, SBIR, etc.)
- **Funding institute:** NIH
- **Fiscal year:** 2024
- **Award amount:** $202,200
- **Award type:** 3
- **Project period:** 1998-02-18 → 2027-01-31

## Primary source

NIH RePORTER: https://reporter.nih.gov/project-details/11159988

## Citation

> US National Institutes of Health, RePORTER application 11159988, Multimodal Federated Learning for Pan-Cancer Detection of Cachexia and Early Detection of Pancreatic Cancer (3P30CA076292-26S3). Retrieved via AI Analytics 2026-05-28 from https://api.ai-analytics.org/grant/nih/11159988. Licensed CC0.

---

*[NIH grants dataset](/datasets/nih-grants) · CC0 1.0*