A hybrid deep learning model that combines inflammatory biomarkers with routine clinical and pathological information has improved postoperative risk prediction in gastric cancer. The evaluation, accepted for publication in BMC Medical Informatics and Decision Making, used a publicly available dataset of 1,330 patients who underwent gastrectomy for stage I–III disease. The framework links tree-based feature modelling with a neural network to identify relationships that may be missed by conventional approaches. It was designed to assess survival after surgery while accounting for tumour characteristics, systemic inflammation and nutritional status, with feature analysis used to clarify which variables most strongly influenced the predictions.
Broader Inputs for Postoperative Risk Assessment
Postoperative prognosis is commonly assessed through tumour stage, differentiation and lymph node involvement, but these features do not fully reflect the patient’s inflammatory and nutritional condition. The model therefore combined standard clinicopathological information with preoperative blood-based measures, including the neutrophil-to-lymphocyte ratio, platelet-to-lymphocyte ratio and a combined platelet and neutrophil score. Age, sex, tumour size, location, grade and stage were also included.
The dataset contained patients with histologically confirmed gastric cancer who had complete clinical, pathological and follow-up records. None had received chemotherapy or radiotherapy before surgery, and patients with recent acute infection or inflammatory disease were excluded. Blood samples were collected shortly before surgery, and overall survival was used as the outcome. Follow-up was scheduled more frequently during the first two years and then at longer intervals.
Must Read: Synthetic Data Supports Cancer Survival Model Transfer
After data checks and preprocessing, records with missing key prognostic or survival information were removed, while categorical and continuous variables were prepared for modelling. The data were then separated into training, validation and test groups while maintaining a similar balance of survival outcomes across each group. This design allowed the framework to be developed and tuned on one portion of the data before being assessed on patients not used during model training, reducing the risk that its reported performance reflected the development sample alone.
Hybrid Architecture Strengthens Predictive Performance
The framework combines three main stages. A gradient-boosted tree component first identifies complex dependencies among clinical and inflammatory variables. A tree-driven encoder then converts the resulting decision paths into a uniform binary representation. A one-dimensional convolutional neural network processes these encoded features and produces a postoperative survival probability or risk score. This structure is intended to preserve the strengths of tree-based modelling while allowing the neural network to learn higher-level patterns from heterogeneous clinical information. Performance was assessed on an independent test set using standard measures of classification and discrimination. The model reached an area under the receiver operating characteristic curve of 0.902 and performed better than the tested conventional machine learning methods and more recent deep learning models for tabular data.
The strongest tree-based comparator and the best-performing advanced deep learning comparator both produced lower discrimination. Repeated training with different random seeds showed little variation in the results, indicating that the model converged consistently under different experimental conditions. The findings suggest that combining tree-based structuring with deep feature learning improved prediction across the available dataset. The binary encoding stage also created a common representation for different types of clinical information, allowing the neural network to process demographic, pathological and inflammatory variables within the same feature space.
Inflammatory Markers Shape Individual Predictions
Feature attribution analysis showed that inflammatory and nutritional measures contributed substantially to the model’s predictions. The neutrophil-to-lymphocyte ratio, C-reactive protein and albumin were among the most influential variables. Higher values for the inflammatory ratio and C-reactive protein were associated with greater predicted risk, while higher albumin levels were linked to lower predicted risk. The analysis also showed that the effect of a given biomarker varied according to the wider clinical profile. Inflammatory measures therefore did not operate as isolated thresholds but interacted with tumour stage, location, age, tumour size and histological grade. Among the inflammation-based measures, the neutrophil-to-lymphocyte ratio showed the strongest relationship with postoperative survival.
Patients with values above three had lower survival, while higher platelet-to-lymphocyte ratios were also associated with greater postoperative mortality. Tumour stage and histological grade remained important, but their prognostic meaning changed when considered alongside inflammatory measures. Some patients with stage II disease and elevated inflammatory ratios had mortality risks similar to those with stage III disease. These patterns support the inclusion of systemic inflammation and nutritional status in postoperative risk assessment, although the model used static variables and did not examine how these measures changed after surgery. The deep learning component also remained only partly interpretable despite the use of feature attribution methods, limiting the transparency of individual predictions.
The hybrid model improved postoperative prognostic classification by combining clinicopathological information with inflammatory and nutritional markers. It outperformed the tested conventional and deep learning comparators and remained stable across repeated training runs. Its predictions reflected both tumour characteristics and the patient’s systemic condition, with inflammatory ratios, C-reactive protein and albumin among the leading contributors. Clinical use remains limited by the dataset size, its restricted institutional base, reliance on static variables and incomplete model transparency. External validation in independent centres, longitudinal data and stronger explanation methods are needed before routine clinical application.
Source: BMC Medical Informatics & Decision Making
Image Credit: iStock
References:
Zhu Q, Chen G, Mao Z et al. (2026) Inflammatory marker-driven deep learning model for postoperative gastric cancer prognosis. BMC Med Inform Decis Mak. https://doi.org/10.1186/s12911-026-03661-4
