Prediction of 5-year overall survival of tongue cancer based machine learning

Data were obtained from the electronic medical records of 224 patients with OTSCC reported at the PLA General Hospital from August 2009 to December 2017, containing 51 clinical features as follows: age, sex, height, weight, body mass index (BMI), hypertension, diabetes, white blood cell count (WBC), neutrophil percentage (N), lymphocyte percentage (L), monocyte percentage (M), platelet count (PLT), lymphocyte-to-monocyte ratio (LMR), platelet-to-lymphocyte ratio (PLR), neutrophil-to-lymphocyte ratio (NLR), systemic inflammation response index (SIRI), hematocrit (Hct), mean cellular hemoglobin concentration (MCHC), average platelet volume (MPV), activated partial thrombin time (APTT), plasma fibrinogen (FIB), hemoglobin (Hb), albumin, glycosylated Hb, targeted therapy, tumor size, tumor location, T-stage, N-stage, positive lymph nodes, histologic grade, OTSCC classification, urinary specific gravity (SG), urinary red blood cell count (RBC), blood urea nitrogen (BUN), serum creatinine (SCR), serum uric acid (SUA), total bilirubin (T-BiL), direct bilirubin (D-BiL), homocysteine (HCY), γ-glutamine transferase (GGT), random blood glucose (RBG), total cholesterol (TC), triglyceride (TG), high-density lipoprotein (HDL), low density lipoprotein (LDL), calcium (Ca), phosphorus (P), serum potassium (K), serum sodium (Na), and bicarbonate.

Inclusion criteria were (1) patients with OTSCC presenting to the PLA General Hospital for the first time; (2) patients with a pathological diagnosis of OTSCC; (3) all patients had complete clinical records and follow-up data. Exclusion criteria comprised (1) patients who had a cold one week before surgery; (2) patients with other tumors; (3) patients receiving anti-tumor treatment before surgery. After applying strict inclusion and exclusion criteria, 224 patients finally met the requirements. The endpoint event of the present study was the overall survival rate (OS). The OS was defined as the interval between the date of surgery and death or the last follow-up. The last follow-up date was 1 April 2022. The flow chart of this study is shown in Fig. 1.

With survival time and survival status as the outcome events, 18 characteristic variables with a significant correlation were selected by Cox proportional hazards model. Then, the top 8 important feature variables were selected from the 18 significantly correlated variables through LGBM, and secondary modeling was conducted through the grid search.

Predictive models were used to construct the 5-year overall survival of OTSCC patients using six machine learning methods, specifically, Logistic Regression, Support Vector machines (SVC), Decision Tree, Random Forest (RF), eXtreme Gradient Boosting (XGB), and Light Gradient Boosting Machine (LGBM).

Logic regression is an algorithm very similar to linear regression, but, in essence, the problem type treated by linear regression is not consistent with logical regression, and linear regression deals with numerical problems, while logical regression belongs to the classification algorithm [9]. Support vector machine is a 2-classification algorithm, which adopts the kernel skill based on mapping the input data to the high-dimensional feature space through the nonlinear transformation to achieve the linear separation of the high-dimensional space [10]. A Decision Tree is an example-based inductive learning algorithm that divides the disordered samples into different branches according to certain rules according to the characteristics of the samples to achieve the purpose of classification or regression [11]. Additionally, XGB, LGBM, and RF are also very commonly used algorithms in machine learning [12‐14].

Using survival time and survival status as outcome events, the final output of the prediction model was defined as the 5-year OS of patients with OTSCC.

In total, 224 OTSCC patients were included in the study. Among them, 150 patients were males (67.0%), and 74 patients (33.0%) were females. There were 136 cases (60.7%) of patients aged < 60 years, and 88 cases (39.3%) of patients aged > 60 years. Tumor size was ≤ 4 and > 4 in 193 (86.2%) and 31 (13.8%) cases, respectively. T-stage was T1, T2, and T3 in 69 (30.8%), 124 (55.4%), and 31 (13.8%) cases, respectively. N-stage was N0, N1, N2, and N3 in 129 (57.6%), 47 (21.0%), 44 (19.6%), and 4 (1.8%) cases, respectively. OTSCC classification was I, II, III, and IV in 48 (21.4%), 69 (30.8%), 12 (5.4%), and 95 (42.4%) cases, respectively. Histologic grade was I, II, and III in 103 (46.0%), 102 (45.5%), and 19 (8.5%) cases, respectively. Lymph nodes were positive in 95 (42.4%) cases and negative in 129 (57.6%) cases. The mean N was 0.58 ± 0.09, and the mean L was 0.32 ± 0.09. The mean SIRI was 1.27 ± 0.92, the mean LMR was 2.22 ± 1.54, the mean PLR was 5.16 ± 76.96, the mean SG was 1.02 ± 0.01, the mean WBC was 6.13 ± 1.86, mean FIB was 3.41 ± 0.91, mean HCY was 14.61 ± 6.50, mean albumin was 41.39 ± 3.75, and mean Na was 142.24 ± 2.67 (Table 1).

Among the 51 clinical features, 18 variables were selected by Cox proportional hazards model, and p-value of each variable was less than 0.05. The 18 variables were age, tumor size, T-stage, N-stage, OTSCC classification, histologic grade, positive lymph nodes, N, L, SIRI, LMR, PLR, SG, WBC, FIB, HCY, albumin, and Na (Fig. 2).

Six machine learning was performed on 18 variables to predict 5-year survival in OTSCC patients. The performance of 6 machine learning models is shown in Table 2. ROC curves under six machine learning are shown in Fig. 3. Random Forest (RF) had the maximum AUC value (AUC = 0.850), and eXtreme Gradient Boosting (XGB) and Light Gradient Boosting Machine (LGBM) had the minimum AUC value (AUC = 0.790).

After the grid search, the Light Gradient Boosting Machine (LGBM) model had the maximum AUC value (Fig. 4a, AUC = 0.851), exceeding the corresponding AUC value of Random Forest (RF) (AUC = 0.850). SHAP explains the results of the LGBM model by calculating the contribution of each variable to the prediction. The importance matrix plot of the LGBM model with 18 feature variables containing significant correlations is shown in Fig. 4b. The 18 feature variables were N-stage, SIRI, age, FIB, LMR, T-stage, N, positive lymph nodes, histologic grade, HCY, Na, WBC, albumin, L, tumor size, OTSCC classification, PLR, and SG.

From these 18 significant correlation variables, the top 8 feature variables were selected for secondary modeling ROC curve (AUC = 0.860, Fig. 4c). The importance matrix map of the LGBM model is shown in Fig. 4d. The top 8 feature variables were N-stage, age, SIRI, positive lymph nodes, FIB, LMR, N, and T-stage.

Figure 5a demonstrates the SHAP summary plot. Each point in each row represents the records of 224 patients with OTSCC under each feature. These features are ranked from the most important to less important order: N-stage, age, SIRI, positive lymph nodes, FIB, LMR, N, and T-stage. The N-stage is the most important feature. The higher the values of the features, the more positive the predictive effect on survival. The lower the value, the lower the contribution is.

Figure 5b shows the SHAP force plot. The predictive value is 0.42. The base value is the mean of the target feature variable across all records. Each band shows the effect of its characteristics in pushing the value of the target feature variable further or closer to the base value. Red stripes indicate their features pushing values to lower values. Blue stripes indicate their features pushing values to lower values. The wider the stripe, the higher the contribution (absolute value). The LMR and the FIB contributed positively to the predicted values. The N-stage is still the most important feature variable because its contribution is the largest (it has the widest strip).

Figure 5c illustrates the SHAP force plot for LGBM. The abscissa represents each patient, and the ordinate represents the SHAP value. The figure shows the SHAP values for the partial characteristics of some patients. Red indicates a positive correlation, and blue indicates a negative correlation.