Development and Validation of a Personalized Survival Prediction Model for Uterine Adenosarcoma: A Population-Based Deep Learning Study

Abstract

Background: The aim was to develop a personalized survival prediction deep learning model for adenosarcoma patients using the surveillance, epidemiology and end results (SEER) database.

Methods: A total of 797 uterine adenosarcoma patients were enrolled in this study. Duplicated and useless variables were excluded, and 15 variables were selected for further analyses, including age, grade, positive lymph nodes or not, marital status, race, tumor extension, stage, and surgery or not. We created our deep survival learning (DSL) model to manipulate the data, which was randomly split into a training set (n = 519, 65%), validation set (n = 143, 18%) and testing set (n = 143, 18%). The Cox proportional hazard (CPH) model was also included comparatively. Finally, personalized survival curves were plotted for randomly selected patients.

Results: The c-index for the CPH model was 0.726, and the Brier score was 0.17. For our deep survival learning model, we achieved a c-index of 0.774 and a Brier score of 0.14 in the external testing set. In addition, the limitations of the traditional staging system were revealed, and a personalized survival prediction system based on our risk scoring grouping was developed.

Conclusions: Our study developed a deep neural network model for adenosarcoma. The performance of this model was superior to that of the traditional Cox proportional hazard model. In addition, a personalized survival prediction system was developed based on our deep survival learning model, which provided more accurate prognostic information for adenosarcoma patients.

Keywords: adenosarcoma; artificial intelligence; database; deep learning; personalized model; survival prediction.

Figure 1

Correlation matrix of 15 selected features. Values in this figure indicated the correlation coefficient of two corresponding variables. Color indicated strength of correlation, in which dark blue indicated strong positive, and dark red indicated strong negative relationships. Diagnosis: year when patients were diagnosed. Diam: diameter of tumor. Lymex: number of excised lymph nodes. Lympo: number of positive lymph nodes. Lymph: positive lymph nodes or not. Spread: extension of tumor. Surg: surgery or not. Surgery: surgery type.

Figure 2

Performance of cox proportional hazard (CPH) model. (A) CPH model has 0.17 of IBS (below 0.25 obviously). (B) CPH model make a median absolute error of 1.615 patients and mean absolute error of 2.223 patients during 12000 days of follow-up time in testing set. (C) Predicted and actual survival curves plotted by CPH model. It made a median absolute error of 13.726 and mean absolute error of 14.626. As we can see from the figure, some spots were plotted outside the confidence intervals.

Figure 3

Values of loss function for DSL model decrease from 1,300 to 762 after 2,000 time of iterations.

Figure 4

Performance of Deep survival learning (DSL) model. (A) In the independent testing set, we achieved 0.14 of brier score using our DSL model. (B) DSL model made a median absolute error of 1.989 patients and mean absolute error of 2.621 patients during 12,000 day of follow-up time in testing set. (C) DSL model made a median absolute error of 3.851 and mean absolute error of 5.632 in survival curve prediction. Nearly all spots of predicted curve lied within confidence intervals of actual curve and the predicted curve was drew similarly to the actual one.

Figure 5

Survival curves for conventional staging system and personalized survival prediction established by DSL model. (A) K-M curve of conventional staging system showed significant difference between stage I from other three stages and inapparent difference between stage II, III and IV. (B) We divided adenosarcoma patients into three stages according to risk factors calculated by our DSL model. Patients with a score of 0-4 were classified in stage I and marked in red color, patients with a score of 4-5.5 in stage II and green color, patients with 5.5-8 score in stage III and blue color.

Figure 6

Personalized survival curve for three randomly selected patients showed apparently diverse results. After six times Repeated selections and validations (A–F), Patient with low risk always has the best survival result, contrasting with patient with high risk resulting in short survival time.