A deep feature fusion model integrating heterogeneous EHR data achieved high-accuracy lung cancer risk prediction with an AUC of 0.978 (95% CI, 0.9723-0.9828), outperforming benchmark algorithms.
Cohort (n=5,257)
No
Does a deep learning-based feature fusion model improve lung cancer risk prediction compared to benchmark algorithms in a health screening population?
A deep learning feature fusion model using heterogeneous EHR data achieved high accuracy for lung cancer risk prediction, outperforming traditional benchmark algorithms.
Estimación del efecto: AUC 0.978 (95% CI 0.9723-0.9828)
Abstract Importance Lung cancer remains the leading cause of cancer-related mortality worldwide. Existing risk prediction models are constrained by their reliance on a limited set of clinical variables or expensive omics data, which hinders their widespread applicability at the population level. Developing a scalable, high-accuracy prediction model using routinely collected health examination data could transform early detection strategies. Objective To develop and validate a deep learning-based feature fusion model that integrates heterogeneous electronic health record (EHR) data for accurate, cost-effective lung cancer risk prediction in a real-world health screening population. Design, Setting, and Participants This retrospective cohort study analyzed data from the Health Management Center of West China Hospital, Sichuan University, spanning January 1, 2010, to December 31, 2022. Participants were stratified by age, sex, and smoking status to ensure representative sampling. Interventions A novel three-tier deep feature fusion architecture was developed: (i) Heterogeneous encoding layer: Multilayer perceptron (MLP) networks processed structured clinical data (demographics, laboratory results, imaging findings), while bidirectional encoder representations from transformers (BERT) encoded unstructured medical narratives; (ii) Hierarchical fusion layer: Multi-scale feature alignment via dimensional expansion concatenation and gated recurrent unit (GRU) networks enabled deep cross-modal interactions; (iii) Attention-based decision layer: Word-level attention mechanisms with weighted pooling dynamically prioritized predictive features to generate risk probability distributions. Main Outcomes and Measures Primary outcomes were discrimination (area under the receiver operating characteristic curve AUC and area under the precision-recall curve PR curve), calibration (Brier score, calibration plots), and clinical utility (decision curve analysis). Model performance was compared against six benchmark algorithms (logistic regression, random forest, XGBoost, support vector machine, AdaBoost, traditional neural networks, and clinical risk scores). Results Among 5,257 participants (age: 46.03 (16 to 96) years; 2997 (57.01%) female; 1226 (23.32%) ever-smokers), the deep fusion model achieved superior performance on the held-out test set (n = 1,314): sensitivity 0.946 (95% CI, 0.931-0.960), specificity 0.984 (95% CI, 0.980-0.989), AUC 0.978 (95% CI, 0.9723-0.9828) and F1-score of 0.896 (0.881-0.910). The model demonstrated excellent calibration and significantly outperformed the best benchmark models. Subgroup analyses confirmed robust performance across genders, smoking status, and age groups. Conclusions and Relevance This deep feature fusion model, leveraging routinely collected heterogeneous EHR data, achieved high-accuracy lung cancer risk prediction with excellent calibration and clinical utility. This abstract is funded by: None
Tang et al. (Fri,) conducted a cohort in Lung cancer risk prediction (n=5,257). Deep learning-based feature fusion model integrating heterogeneous EHR data vs. Six benchmark algorithms (logistic regression, random forest, XGBoost, support vector machine, AdaBoost, traditional neural networks, and clinical risk scores) was evaluated on Discrimination (AUC and PR curve), calibration, and clinical utility for lung cancer risk prediction (AUC 0.978, 95% CI 0.9723-0.9828). A deep feature fusion model integrating heterogeneous EHR data achieved high-accuracy lung cancer risk prediction with an AUC of 0.978 (95% CI, 0.9723-0.9828), outperforming benchmark algorithms.