Machine Learning Layer Boosts 14-Gene Oral Cancer Test Across Diverse Populations

Researchers have combined a rapid molecular test for oral cancer with machine learning in a way that could make the test easier to deploy in new countries and clinics. In a study published in PLOS Computational Biology, a team from the University of Hertfordshire, Curenetics Ltd, Guy’s and St Thomas’ NHS Foundation Trust and Queen Mary University of London report that wrapping the quantitative Malignancy Index Diagnostic System, or qMIDS, inside machine learning classifiers improved diagnostic performance across four patient cohorts from the United Kingdom, India and China, with a design called late fusion delivering the most consistent gains.

The qMIDS test itself is built around a strikingly small tissue sample. Where a conventional surgical biopsy typically removes more than 5 to 10 millimetres of tissue and may require general anaesthesia, qMIDS needs only a 1 to 2 millimetre punch biopsy. The tissue is digested into a lysate, mRNA is purified using Dynabeads, and a one-step reverse transcription quantitative polymerase chain reaction measures 14 target genes plus two reference genes. The whole workflow takes less than 90 minutes from tissue to result, compared with more than a week for histopathology, and a malignancy index threshold of MI = 4, selected empirically, decides whether a sample is called cancerous.

The 14 target genes, including HOXA7, CENPA, NEK2, DNMT1, INHBA, FOXM1, TOP2A, BIRC5, MMP13, CXCL8, NR3C1, IVL, CBX7 and S100A16, were originally selected from clinical tissue samples in the UK and Norway and later validated in samples from the UK, China and India. They sit within a transcriptional program governed by FOXM1, an oncogenic transcription factor that has been shown to be among the top upregulated genes across 39 cancer types and a major predictor of poor prognosis. The two reference genes, YAP1 and POLR2A, were chosen after a systematic screen of candidate control genes in normal oral keratinocytes, mucosa, dysplasias and head and neck squamous cell carcinoma tissues, in which they proved consistently and stably expressed.

The malignancy index itself is a fixed, closed-form formula. For each of the 14 genes, the sample’s normalised expression is compared with a reference median derived from eight independent healthy primary normal human oral keratinocytes via a log2 ratio. These per-gene deviations are summed, then scaled by a term built from the first quartile, third quartile and maximum of the deviation distribution, so that a sample in which a few genes are strongly dysregulated is treated differently from one in which all genes shift moderately and uniformly. The product is linearly rescaled to give the final MI score, which is compared against the single threshold.

The new study tested two ways of fusing this index with machine learning. In early fusion, the MI score is simply concatenated with the 14 gene-expression values and treated as a 15th feature, eligible for selection or discard alongside the genes by feature selection algorithms, here mRMR and SVM-RFE, before classification. In late fusion, a base classifier trained only on the gene features produces class probabilities, which are then concatenated with the MI score and fed to a meta-classifier. To avoid information leakage, the probabilities used to train the meta-classifier were generated by proper stacking, with out-of-fold predictions made while each training sample was held out of the base classifier’s own fold.

The evaluation covered four cohorts, with the Indian data split into North India and South India subsets, and used ten-fold cross-validation with nested hyperparameter tuning, so the outer test fold was never used for feature selection or model selection. Ten classifiers were tried, including support vector machines with three kernels, random forests, k-nearest neighbours, logistic regression, AdaBoost, gradient boosting, extra trees, linear discriminant analysis and XGBoost. Against the qMIDS-alone baseline, late fusion improved AUC by 0.47, 2.83, 6.07 and 4.40 percentage points in the UK, IN-N, IN-S and CN cohorts respectively, and improved F1 by 1.73 percentage points in IN-S and 6.29 in CN while remaining within 0.7 points of qMIDS alone in the UK and IN-N. Relative to machine learning without fusion, late fusion improved F1 by up to 10.74 and AUC by up to 6.8 percentage points, both in the IN-S cohort. The best late fusion models were LDA in the UK, XGBoost in IN-N, SVM in IN-S and extra trees in CN.

Because some cohorts were small, the authors also probed stability and significance. Standard deviations and 95 percent confidence intervals of AUC across the ten folds were narrow relative to the reported gains, and learning curves showed converging rather than diverging training and validation performance. Paired significance testing against the qMIDS-alone baseline reached conventional significance, at p below 0.05, for the early fusion model in the UK and IN-S cohorts and the late fusion model in the CN cohort; the remaining comparisons showed consistent numerical improvements that did not reach significance at the available sample sizes. The team also compared the two architectures directly across 140 matched classifier, feature-selection and feature-count configurations per cohort. Late fusion was never significantly worse than early fusion in the IN-N, IN-S and CN cohorts, and significantly better in 41, 80 and 68 configurations respectively, while the UK cohort was more balanced at 39 versus 29. Yet at the single best-performing configuration, the two architectures were statistically indistinguishable in IN-N, IN-S and CN, and early fusion was significantly better in the UK.

The most practically important result came from inter-dataset analysis, in which models were trained on one population and tested on another. Performance drops were generally under 2 percent across most train and test combinations: the best inter-dataset result for the UK cohort was 1.17 percentage points lower in F1 and 1.09 lower in AUC than within-cohort testing, while IN-N and IN-S actually gained up to 1.06 and 2.06 F1 points respectively, and CN lost 0.77 F1 and 1.47 AUC points. In a direct matched comparison across 12 train-test combinations and 120 configurations each, late fusion significantly outperformed early fusion in 397 of 1,440 comparisons, and early fusion never significantly outperformed late fusion. The authors interpret occasional raw-AUC advantages for early fusion under population shift as reflecting its ability to learn direct interactions between the MI score and gene features when distributions differ.

Biologically, the team argues that the gains reveal something about the FOXM1 program itself: the diagnostic information carried jointly by pairs and higher-order combinations of the 14 genes exceeds what a single linear index can represent, so a learned meta-classifier recovers residual signal without altering the gene panel. The largest gains appeared in the cohort with the most heterogeneous case mix and smallest sample size, IN-S, where a threshold calibrated on other populations is most likely to be miscalibrated. Improved sensitivity relative to the fixed-threshold baseline, from 82.92 to 93.33 percent in IN-S and from 91.40 to 94.86 percent in CN, may reduce false-negative misclassification and support earlier intervention.

The authors are careful about limits. Cohorts ranged from 33 to 309 samples, with the South Indian cohort at 33 and the Chinese cohort at 35, so several three-significant-figure improvements fall within their confidence intervals and should be read as preliminary. Nested cross-validation and learning curves are not substitutes for prospective, independently collected validation cohorts, and the authors recommend a population-specific validation step with a small locally representative sample before any deployment in a new population, laboratory or sample-handling protocol. They also note that no cross-cohort batch correction was applied, a deliberate choice to preserve genuine between-population signal, and that retrospective, non-batch-corrected data make it hard to fully disentangle technical heterogeneity from biology. With those caveats, the study offers a template the authors call interpretable index plus adaptive machine learning layer, which they suggest could be extended to other single-index molecular diagnostics and, given FOXM1’s breadth across cancers, potentially beyond the oral cavity, pending independent validation.

Subject of Research: Fusion of the qMIDS 14-gene malignancy index with machine learning classifiers for oral cancer diagnosis from RT-qPCR gene expression data

Article Title: Early vs late fusion of qMIDS with machine learning based classifiers for oral cancer diagnosis from gene expression data

Article References: Ben Ali, F., Mporas, I., Balaji, V., Ibaranyanakaye, D., Carey, B., Teh, M.-T., & Adeleke, S. (2026). Early vs late fusion of qMIDS with machine learning based classifiers for oral cancer diagnosis from gene expression data. PLOS Computational Biology, 22(10), e1014888. https://doi.org/10.1371/journal.pcbi.1014888

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014888

Keywords: oral cancer, qMIDS, machine learning, late fusion, early fusion, gene expression, RT-qPCR, FOXM1, malignancy index, cross-population validation, OSCC, diagnostic biomarkers