Explanation-Based Selective Referral in Bimodal Lung Cancer Risk Prediction Using Clinical and Low-Dose CT Data

Principal Investigator

Name
Abdul Sadka

Institution
Aston University

Position Title
Professor and Head of Institute

Email
a.sadka@aston.ac.uk

About this CDAS Project

Study
NLST (Learn more about this study)

Project ID
NLST-1529

Initial CDAS Request Approval
Aug 24, 2026

Title
Explanation-Based Selective Referral in Bimodal Lung Cancer Risk Prediction Using Clinical and Low-Dose CT Data

Summary
Machine learning models for lung nodule malignancy prediction increasingly combine low-dose CT imaging with patient clinical data. While such models often achieve strong aggregate performance, they provide limited insight into when an individual prediction should be trusted. Conventional uncertainty estimates are known to fail in a specific and clinically important way: some incorrect predictions are made with high confidence, and these are precisely the cases least likely to be flagged for human review.

This doctoral project investigates whether disagreement between the clinical and imaging evidence used by a bimodal model provides error-detection information beyond conventional predictive uncertainty. The central hypothesis is that when the clinical profile and the CT findings push a prediction in opposing directions, the model is more likely to be wrong, and such cases warrant radiologist review.

The work has three components. First, a bimodal malignancy prediction model will be developed, combining participant clinical variables with quantitative imaging features derived from segmented pulmonary nodules. Second, modality-specific explanations will be generated — feature attribution for the clinical branch and spatial attribution for the imaging branch — and evaluated for faithfulness using perturbation-based deletion and insertion analyses, ensuring that explanations reflect model behaviour before their disagreement is interpreted. Third, a signed measure of directional conflict between modality contributions will be computed at the model-output level and tested as a selective-referral signal against established baselines including predictive entropy, Monte Carlo dropout, and conformal prediction.

The cohort will follow the established NLST nodule-positive control design used in prior work from the Moffitt Cancer Center group, comparing participants with screen-detected nodules who were subsequently diagnosed with lung cancer against participants with nodules who were not, matched on nodule size and relevant participant characteristics.

Data requested and justification. Imaging will be obtained from the NLST collection hosted by The Cancer Imaging Archive. The publicly distributable clinical subset available through TCIA contains only four usable participant-level predictors: age, sex, race, and binary smoking status. This is insufficient to construct a clinical modality capable of contributing meaningful independent evidence alongside CT imaging, which is a precondition for the study. Access is therefore requested to the full participant dataset, specifically quantitative smoking history including pack-years and time since cessation, medical history including chronic obstructive pulmonary disease and other comorbidities, family history of cancer, and occupational exposure variables. Screening and abnormality tables are also requested to support nodule-level cohort construction.

All data will be held on secured institutional systems at Aston University, used solely for the purposes described, and not shared beyond the approved research team. No attempt will be made to re-identify participants. Results will be reported in aggregate in a doctoral thesis and peer-reviewed publications, with appropriate citation of the NLST and acknowledgement of the National Cancer Institute.

Aims

To construct a nodule-level cohort from NLST comprising participants with screen-detected pulmonary nodules subsequently diagnosed with lung cancer and matched nodule-positive participants who were not, following the established positive-control design used in prior NLST radiomics research.

To develop and evaluate a bimodal malignancy prediction model combining participant clinical variables with quantitative imaging features extracted from segmented pulmonary nodules on low-dose CT.

To determine whether the clinical modality contributes independent predictive information beyond CT imaging alone, by comparing clinical-only, imaging-only, and fused models.

To generate modality-specific explanations for each branch of the model and assess their faithfulness using perturbation-based deletion and insertion analyses.

To define a signed measure of directional conflict between clinical and imaging evidence at the model-output level, distinguishing it from modality contribution imbalance and cross-modal interaction.

To test whether faithfulness-qualified directional conflict identifies model errors beyond established uncertainty baselines including predictive entropy, Monte Carlo dropout, and conformal prediction.

To characterise the conditions under which explanation conflict is and is not informative, including analysis of confidently incorrect predictions that conventional uncertainty methods fail to flag.

To evaluate the resulting conflict measure as a selective-referral signal using risk–coverage analysis, quantifying error reduction achievable at clinically plausible referral rates

Collaborators

Abdul Sadka Aston University
saleem khan Aston University