AI Model Using EHR Data Predicts 5-Year Survival in Colorectal Cancer

Transforming electronic health record (EHR) data into image matrices enables powerful deep learning models to predict 5-year survival outcomes in colorectal cancer (CRC) with improved accuracy, according to one study. Using the Visual Geometry Group (VGG16) architecture, researchers achieved up to 78% accuracy and strong specificity, while explainable artificial intelligence (AI) techniques highlighted clinically relevant features driving predictions.

The findings are published in JMIR Medical Informatics.

A novel image-based deep learning approach achieves high accuracy and interpretability, offering potential for clinical decision support. | Image credit: Ahmet Aglamaz - stock.adobe.com

“In this study, we developed a model to predict the survival period of colorectal cancer using [EHR] data and investigated which variables contributed to the prediction,” wrote the researchers of the study. “In particular, we improved the performance of the model through an innovative approach to convert tabular medical data into image data. The results of the study showed that the VGG16 model achieves the best performance, which suggests a new methodology for developing a CDSS for patients with colorectal cancer in clinical settings.”

A previous study found that machine learning models integrating clinical and sociodemographic variables could reliably predict 5-year postoperative survival in patients with stage III CRC.² Key prognostic factors included age, lymph node ratio, chemotherapy status, and tumor stage, along with marital status, tumor location, and histological type.

The current study analyzed anonymized EHRs from 3321 patients with CRC, incorporating demographic data, tumor characteristics, laboratory results, treatment modalities, and follow-up outcomes.¹ The researchers utilized the Image Generator for Health Tabular Data, which transformed structured clinical variables into 2D image matrices. Patients were stratified into colon and rectal cancer subgroups to reflect biological and prognostic differences. Three predictive models were developed and compared: a conventional artificial neural network (ANN), a convolutional neural network (CNN), and a transfer learning–based Visual Geometry Group (VGG16) model.

Among the 3 models tested, the transfer learning–based VGG16 architecture demonstrated the strongest predictive performance for 5-year survival in CRC. For colon cancer, VGG16 achieved an accuracy of 78.44% with a notably high specificity of 89.55%. For rectal cancer, the model reached an accuracy of 74.83% and a specificity of 87.9%.

In comparison, CNN showed lower overall accuracy and poor specificity, limiting its clinical applicability. VGG16 also maintained a better balance between sensitivity and specificity, making it more robust in handling moderate class imbalance. Gradient-weighted class activation mapping (Grad-CAM) visualizations further validated the model by highlighting clinically relevant prognostic factors such as age, gender, smoking history, American Society of Anesthesiologists physical status grade, liver disease, pulmonary disease, and initial carcinoembryonic antigen levels.

However, the researchers noted several limitations. The dataset was drawn from a single institution with a relatively small sample size, which may have limited generalizability. Additionally, the image matrix used to represent 25 clinical variables was arbitrarily defined, in which data-driven layouts may better reflect clinical relationships. Although Grad-CAM provided interpretability, the researchers noted that real-world use will require standardized EHR integration, interoperability, and external validation to ensure predictions are accurate, actionable, and clinically trusted.

Despite these limitations, the researchers believe the study suggests the VGG16 approach shows promise as a future clinical decision support tool for CRC care.

“Deep learning models performed important clinical predictions on whether the survival period of patients with colorectal cancer who underwent surgery is more than 5 years,” wrote the researchers. “If the survival period of a patient with cancer is more than 5 years, it indicates that the patient’s prognosis is good and the risk of cancer recurrence is low, which can provide clinically helpful information. Furthermore, it shows that doctors can use this as a reference to understand the patient’s condition and support them in making better decisions.”

References

1. Oh SH, Lee Y, Baek JH, et al. Learning and image generator health tabular data (IGHT) for predicting overall survival in patients with colorectal cancer: retrospective study. JMIR Med Inform. 2025;13:e75022. doi:10.2196/75022

2. Steinzor P. Machine learning predicts 5-year survival in stage III CRC. AJMC®. July 29, 2025. Accessed August 19, 2025. https://www.ajmc.com/view/machine-learning-predicts-5-year-survival-in-stage-iii-crc