Developing multimodal cervical cancer risk assessment and prediction model based on LMIC hospital patient card sheets and histopathological images

Abstract

Cervical cancer remains a significant global health burden, particularly in low- and middle-income countries (LMICs) where access to early diagnostic tools is limited. In Ethiopia, cervical cancer diagnosis often relies on manual interpretation of biopsies, which can be time-consuming and subjective. This study aims to develop a multimodal machine learning model that integrates histopathological images and associated patient clinical records to improve cervical cancer risk prediction and biopsy detection. The dataset comprises 404 biopsy images and corresponding clinical records from 499 patients, collected at Jimma Medical Center. The preprocessing of histopathological images and clinical records involved image enhancement, data augmentation, imputation of missing values, and class balancing techniques. Subsequently, (I) a pre-trained convolutional neural network deep learning (VGG16) model was applied on the histopathological dataset, (II) a Random Forest classifier was trained on the patient clinical records, and (III) a late fusion strategy was employed to integrate the outputs of both classifiers for multimodal analysis. Recursive Feature Elimination was used to identify key predictive factors from the patient data, and the model’s performance was thoroughly validated using accuracy, AUC-ROC curves, and confusion matrices, ensuring reliability across all classes. As a result, convolutional neural networks and Random Forest classifiers achieved accuracies of 91% and 96%, respectively. The integrated multimodal model achieved 92% accuracy, demonstrating enhanced robustness and clinical relevance by combining complementary data sources. These findings suggest that multimodal approaches hold promise for improving cervical cancer diagnostics in resource-limited settings. Future work will focus on validating the model with diverse datasets and integrating it into clinical workflows to support healthcare providers in LMICs.

Keywords: Cervical cancer; Deep learning; LMICs; Machine learning; Multimodal model; Predictive model; Risk analysis.

Fig. 1

Proposed methodology

Fig. 2

Pipeline for multimodal processing model

Fig. 3

Histopathological processing pipeline

Fig. 4

Patient card-sheet classification architecture

Fig. 5

Sample histopathology images (subtle pathological features)

Fig. 6

Class visualization before and after data augmentation

Fig. 7

Patient card-sheet class distribution after resampling

Fig. 8

Mean age of the women facing the risk of cervical cancer

Fig. 9

ROC curve for multi-class classification of training and testing sets

Fig. 10

Training and validation accuracy

Fig. 11

Training and validation loss

Fig. 12

Roc curves for VGG16

Fig. 13

VGG16 confusion matrix

Fig. 14

ResNet50 model training and validation accuracy

Fig. 15

ResNet50 model training and validation loss

Fig. 16

ResNet50 ROC curve

Fig. 17

ResNet50 confusion matrix

Fig. 18

Confusion matrix for RF model

Fig. 19

ROC curve for RF model

Fig. 20

Confusion matrix curve for the decision tree model

Fig. 21

ROC curve for the decision tree model