# A Quantitative Risk Model for Predicting Outcome and Identifying Structural Biomarkers of Treatment Targets in Oral Cancer on a Large Multi-Center Patient Cohort

> **NIH NIH R01** · STATE UNIVERSITY OF NEW YORK AT BUFFALO · 2020 · $382,900

## Abstract

Post-resection prognostication for oral cavity cancers (OCC) is qualitative and potentially ambiguous. A
significant subset (25-37%) of Stage I/II patients still develop local recurrence after treatment with surgery alone.
The long-term goal of this proposal will be to create a Quantitative Risk Model (QRM) using machine learning
and artificial intelligence to predict recurrence risk for Stage I/II patients using image-based biomarkers of
aggression. The objective is to develop and validate state-of-the-art systems for biomarker imaging,
quantification, and modeling to accurately predict risk of recurrence in cancer patients based on image analytics.
The central hypothesis is that a quantitative, artificial intelligence approach to pathology will result in significantly
greater prognostic value compared with manual microscope-based analysis. The rationale for this work is that
tumor aggression can be predicted from patterns present in pathology images, given the existence of histological
risk models that have been clinically validated in the past; however, these risk models are not in widespread use
because they are less accurate, robust, and transportable to the larger community of pathologists. This proposal
will test the central hypothesis through three specific aims: (1) Develop an analysis pipeline that can accurately
predict recurrence risk for Stage I/II OCC patients and identify treatment targets (e.g. adaptive local immune
response and angiogenesis); (2) Demonstrate robust performance across a multi-site data cohort collected from
seven national and international centers; and (3) Distil the results of QRM analysis to synoptic pathology
reporting, demonstrating the ability of QRM to interface with standard clinical reporting tools. The innovation for
addressing these aims comes from a unique application of active learning for training artificial intelligence to
recognize tissue structures, new features for quantifying tissue architecture based on the interface between
tumor and host, and a novel approach for large cross-site validation. Moreover, this proposal develops a unique
mapping between computational pathology and commonly-used synoptic reporting variables, enabling rapid
uptake of this work into existing clinical workflows. This research is significant because it provides personalized
outcome predictions for a niche group of undertreated patients with limited options and can serve as the
foundation for designing future clinical trials through identification of treatment targets. Multi-site training and
evaluation, combined with AI-to-report mapping, will be broadly applicable to a large group of computational
approaches, bridging the gap between engineering research labs and clinical application. The expected outcome
of this work is a trained model for predicting Stage I/II OCC recurrence, identification of treatment targets, and
mapping to synoptic reports, as well as a broadly-applicable workflow for the broader computational pa...

## Key facts

- **NIH application ID:** 9974099
- **Project number:** 1R01DE028741-01A1
- **Recipient organization:** STATE UNIVERSITY OF NEW YORK AT BUFFALO
- **Principal Investigator:** Scott Doyle
- **Activity code:** R01 (R01, R21, SBIR, etc.)
- **Funding institute:** NIH
- **Fiscal year:** 2020
- **Award amount:** $382,900
- **Award type:** 1
- **Project period:** 2020-04-23 → 2025-03-31

## Primary source

NIH RePORTER: https://reporter.nih.gov/project-details/9974099

## Citation

> US National Institutes of Health, RePORTER application 9974099, A Quantitative Risk Model for Predicting Outcome and Identifying Structural Biomarkers of Treatment Targets in Oral Cancer on a Large Multi-Center Patient Cohort (1R01DE028741-01A1). Retrieved via AI Analytics 2026-06-11 from https://api.ai-analytics.org/grant/nih/9974099. Licensed CC0.

---

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