Adaptive Diagnostic Testing Algorithms for Design of Cost-sensitive Composite Biomarker Tests

Principal Investigator

Name
Brian Denton

Degrees
Ph.D.

Institution
The University of Michigan

Position Title
Professor

Email
btdenton@umich.edu

About this CDAS Project

Study
PLCO (Learn more about this study)

Project ID
PLCO-2057

Initial CDAS Request Approval
Jul 14, 2026

Title
Adaptive Diagnostic Testing Algorithms for Design of Cost-sensitive Composite Biomarker Tests

Summary
Summary: The goal of this project is a methodological evaluation of optimal test sequencing in the presence of correlated test results and heterogeneous test burdens. Recent discoveries of many new biomarkers are helping clinicians identify early signs of cancer. At the same time, these advances make clinical decision-making difficult because of the potential for false-positive and false-negative results. The former results in unnecessary referral of patients for expensive and invasive tests, such as biopsies and radiological imaging; the latter causes a disease to go undetected. Since no single biomarker is satisfactory on its own, attention is turning to ways to combine biomarkers into composite tests. There are many open questions about which tests to use, when to use them, and how to combine them to screen for diseases in a way that balances the benefits of early detection with the potential costs of testing (e.g., monetary costs, delayed diagnosis, false-negative, and false-positive test results). The most natural approach is to use a “one-shot” method that collects all biomarker tests simultaneously and subsequently predicts health outcomes from the multidimensional data. However, this can be prohibitively expensive, especially for proprietary tests, unnecessarily burdensome to patients, and time-consuming.
This project seeks to apply our optimization modeling framework to Ovarian Cancer to illustrate potential benefits in terms of diagnostic accuracy and cost that can be achieved by selectively using tests identified adaptively, using prior test results to guide additional testing. We propose to use de-identified patient-level data, including CA-125, HE4, ultrasound features, diagnostic procedures, and final pathology/cancer outcomes, for Cervical Cancer. Specifically, we have developed mathematical optimization models for dynamic biomarker selection and sequencing to optimize decisions given the following criteria: i) sensitivity and specificity of the overall composite test design; ii) total cost of implementing a composite test based on the number and type of tests used for diagnosis. The problems we propose to address are challenging because of their stochastic and combinatorial nature, as well as the partially observable nature of latent cancer. We have developed and tested our model on synthetic datasets to demonstrate the computational feasibility of algorithms for solving these models. Our next step is to secure a real dataset of patients who underwent multiple biomarker tests, along with a ground-truth gold-standard test, so we can validate our model on real-world data.

Our analysis plan involves splitting the data 70/30% to create training/testing datasets. Our optimization model will be used to design an optimal adaptive testing protocol, which will be applied to out-of-sample testing data to evaluate the predictive performance of the optimized test design and the estimated testing costs. We hypothesize that optimally designed adaptive tests will achieve near-optimal discrimination but at significantly reduced total cost. We will use bootstrapping to estimate statistical confidence intervals on the performance measures. We will also conduct numerical experiments that measure the computational complexity of our proposed algorithms.

Aims

Aim 1. Curate and statistically characterize an ovarian cancer biomarker dataset for adaptive diagnostic modeling. We will obtain a de-identified dataset with multiple biomarker results per patient and gold-standard disease status, preprocess the data, estimate biomarker correlations and diagnostic accuracy, and assign test cost or burden measures.

Aim 2. Optimize adaptive biomarker testing protocols for ovarian cancer diagnosis. We will use training data to optimize the design of sequential testing policies that determine which biomarker to order next, conditional on prior test results, while balancing sensitivity, specificity, false-positive and false-negative consequences, and total testing cost.

Aim 3. Validate adaptive testing protocols against one-shot biomarker panels and benchmark strategies. We will evaluate the optimized protocols on held-out test data using diagnostic accuracy, expected cost, number of tests used, and bootstrap confidence intervals, and we will assess computational scalability through numerical experiments.

Our central hypothesis is that ovarian cancer biomarker tests exhibit an exploitable correlation structure that enables adaptive testing protocols to achieve diagnostic performance close to that of comprehensive biomarker panels and diagnostic procedures while using fewer tests and reducing expected testing costs. We anticipate that successful completion of this project will provide a validated decision-analytic framework for designing cost-conscious adaptive biomarker testing protocols. The resulting methodology could support more efficient cancer screening and diagnostic pathways by reducing unnecessary testing, limiting false-positive referrals, and preserving the benefits of early detection.

Collaborators

Brian Denton The University of Michigan
Viswanath Nagarajan The University of Michigan
Blake Harris The University of Michigan