Development and External Validation of Machine Learning Models to Predict Postoperative Transitional Cell Carcinoma Mortality: A Retrospective Study

doi:10.21203/rs.3.rs-4874784/v1

Purpose

Transitional cell carcinoma (TCC) of the bladder, also known as urothelial carcinoma, is the most common type of bladder cancer. The primary objective of this study was to develop machine learning (ML) models to predict the survival outcomes of bladder cancer patients.

Methods

This study utilized the SEER dataset to develop an ML model for predicting postoperative disease-specific overall survival in TCC patients. This retrospective study included patients diagnosed between 2010 and 2015 for model training and validation and patients diagnosed after 2018 for external validation. The performance was assessed via sensitivity, specificity, accuracy, likelihood ratios, predictive values, ROC-AUC, and the Brier score. SHAP values and partial dependence plots were used to evaluate feature importance and interactions.

Results

The deep neural network model slightly outperformed the other models (linear regression and gradient boosting) across all the datasets. For the external validation dataset, the model had a sensitivity of 69.40% (95% CI: 67.72–71.05%), specificity of 85.32% (95% CI: 84.91–85.72%), and accuracy of 83.84% (95% CI: 83.43–84.24%). The ROC-AUC was 0.8758 (95% CI: 0.8698 to 0.8816). The model demonstrated a negative likelihood ratio of 0.36 (95% CI: 0.34 to 0.38) and a positive likelihood ratio of 4.73 (95% CI: 4.56 to 4.90). The use of external validation indicated minimal overfitting. The key features influencing mortality included tumor stage, age, tumor size, and grade.

Conclusion

Machine learning models, particularly deep neural networks, can accurately predict mortality in individuals with TCC of the bladder.

Machine Learning

Urothelial Carcinoma

Artificial Intelligence

Cancer

Survival Analysis

Transitional cell carcinoma (TCC) of the bladder, also known as urothelial carcinoma, is the most common type of bladder cancer, accounting for approximately 90% of cases [1]. Epidemiologically, bladder cancer is the tenth most common cancer worldwide, with approximately 573,000 new cases and 213,000 deaths annually [2].

Risk factors for TCC include smoking, which is the most significant factor; occupational exposure to certain chemicals (such as aromatic amines and polycyclic aromatic hydrocarbons); chronic bladder inflammation; and certain genetic predispositions [3]. The prognosis of bladder cancer varies significantly on the basis of the stage and grade at diagnosis [4]. Nonmuscle invasive bladder cancer (NMIBC) has a relatively favorable prognosis but is characterized by frequent recurrence, whereas muscle invasive bladder cancer (MIBC) has a worse prognosis due to higher rates of progression and metastasis [5].

The standard treatment for TCC involves a combination of surgical interventions, such as transurethral resection of bladder tumor (TURBT) for NMIBC and radical cystectomy for MIBC, along with chemotherapy, radiation therapy, and immunotherapy. The choice of treatment depends on the stage and grade of the tumor, as well as the patient's overall health and preferences [6, 7].

In recent years, the application of machine learning (ML) in medicine has garnered significant attention because of its potential to increase diagnostic accuracy, prognostication, and personalized treatment planning [8]. ML algorithms can analyze large datasets to uncover patterns and insights that might be overlooked by traditional statistical methods [9]. In oncology, ML has been used for various purposes, including predicting cancer recurrence, patient survival, and treatment response [10, 11].

The SEER dataset [12] provides a rich resource for developing and validating ML models. This study leverages the SEER dataset to create a robust ML model aimed at predicting postoperative disease-specific overall survival in patients with TCC of the bladder. The study design is retrospective, utilizing data from patients diagnosed between 2010 and 2015 for model training and internal validation and data from patients diagnosed after 2018 for external validation. This approach ensures that the model is both accurate and generalizable, mitigating the risks of overfitting and selection bias.

The primary objective of this study was to develop an ML model using demographic, clinical, and treatment data from the SEER database to predict the survival outcomes of bladder cancer patients.

Study Design

The SEER dataset comprises the cancer data of nearly 30% of the USA population [13], making the data representative and minimizing the chances of overfitting and selection bias. The study design was retrospective. The following information was extracted for each patient from the SEER dataset: age (expressed in five-year intervals), sex (male or female), and race (white, black, Asian or Pacific Islander, American Indian/Alaskan Native). Treatment data included whether the patient underwent chemotherapy or radiation therapy. The disease extent was expressed as tumor size in millimeters, tumor stage, tumor node metastasis (TNM) stage according to the 7th edition of the AJCC staging system, and tumor grade (I to IV). Only patients who underwent cancer-directed surgery were included in this study. Patient survival status was considered either alive or dead. The cause of death was transitional cell carcinoma of the bladder in all patients. Therefore, the postoperative disease-specific overall survival of patients with TCC of the bladder was assessed in this study. All patients were diagnosed between 2010 and 2015.

Another dataset with the same features from patients diagnosed after 2018 from the SEER dataset was used for external validation. To make this dataset compatible with the older dataset, the grade feature was engineered. The 2018 dataset was divided into high-grade and low-grade grades. To make the grades equivalent, high grades were randomly converted to grades III and IV for each case with a 1:1 chance. The same approach was used for low-grade disease, converting it to Grade II and Grade I.

The age of each individual patient was considered the midpoint of the five-year interval. Tumor size and age were normalized via a standard scale. All other variables were one-hot encoded, as they were categorical. For numerical variables, the mean and standard deviation were reported, and for categorical variables, the count and percentage were provided.

Patients with missing data for any variable were excluded from the study. This exclusion was justified because of the large size of the dataset, which ensured enough instances to learn the general pattern in the data.

The large language model GPT-4o OpenAI was used for grammar and general language refinements. All changes were supervised by authors, and they take all the responsibility for it.

Machine learning

The 2010-2015 dataset was split into a 14:3:3 ratio for training, testing, and internal validation. The models were trained on the training data and repeatedly evaluated on the test data to identify the best hyperparameters and model architecture. Once the optimal model was determined, it was tested only once on the internal validation dataset to ensure generalization to real-world data and mitigate overfitting. The model with the best performance was then tested on the 2018+ dataset for external validation.

The models were evaluated via the following metrics: sensitivity, specificity, accuracy, positive and negative likelihood ratios, positive and negative predictive values, and area under the receiver operating curve (ROC-AUC). The Brier score was used to assess model calibration, indicating the accuracy of the predicted probabilities.

A gradient boosting model, linear regression, and a deep neural network were the selected algorithms. These algorithms represent ensemble learning, a classic statistical model, and a novel widely adopted accurate model, respectively. This selection ensures a comprehensive approach to the data analysis.

ROC-AUC curves and detection error trade-offs were used to compare model performance on the test and internal validation datasets. SHapley Additive exPlanations (SHAP) values [14] and partial dependence plots were employed to determine feature importance, nonlinearity, and interactions.

Finally, a model was trained using all available data, and a desktop application with a graphical user interface (GUI) was developed to facilitate the use of the model in daily clinical practice.

Demographics

2010-2015 dataset:

The dataset included 24,481 cases, with a mean patient age of 67.95 years (standard deviation 11.21 years). The average tumor size was 34.79 mm (standard deviation of 20.74 mm), ranging from 1 mm to 150 mm. The sex distribution indicated a male predominance: 18,299 males (74.8%) and 6,182 females (25.2%). The racial composition was primarily white (21,842 cases, 89.2%), followed by black (1,562 cases, 6.4%), Asian (877 cases, 3.6%), and other races (200 cases, 0.8%).

The most common tumor grade was grade IV, which was observed in 11,425 patients (46.7%). Tumor stage "0a" was most common (10,588 patients, 43.3%). The Ta classification was the most prevalent tumor characteristic and was present in 10,588 patients (43.3%). In the N classification, N0 was predominant (22,727 cases, 92.8%), and in the M classification, M0 was observed in 23,668 cases (96.7%), with the remainder classified as M1.

Radiation treatment was administered to 1,361 patients (5.6%), whereas 23,120 patients (94.4%) did not receive radiation. The chemotherapy data revealed that 15,800 patients (64.6%) did not undergo chemotherapy, whereas 8,681 patients (35.4%) did.

Patient outcomes revealed that 17,256 patients (70.5%) were alive, whereas 7,225 patients (29.5%) died due to TCC.

2018+ dataset:

The dataset included 32,348 cases, with a mean patient age of 70.61 years (standard deviation of 10.73 years). The average tumor size was 36.62 mm (standard deviation 48.50 mm), ranging from 0 mm to 990 mm. The sex distribution indicated a male predominance: 24,657 males (76.2%) and 7,691 females (23.8%). The racial composition was primarily white (28,777 cases, 89.0%), followed by Asian or Pacific Islander (1,743 cases, 5.4%), black (1,701 cases, 5.3%), and American Indian/Alaskan Native (127 cases, 0.4%).

The most common tumor grade was Grade IV (10,788 cases, 33.4%), followed by Grade III (10,553 cases, 32.6%), Grade I (5,559 cases, 17.2%), and Grade II (5,448 cases, 16.8%). Tumor stage "0a" was most frequently noted (15,930 cases, 49.2%), followed by Stage I (8,016 cases, 24.8%), Stage II (4,140 cases, 12.8%), Stage III (2,330 cases, 7.2%), Stage 0 (1,054 cases, 3.3%), and Stage IV (878 cases, 2.7%).

In the T classification, Ta was the most prevalent (15,930 cases, 49.2%), followed by T1 (8,208 cases, 25.4%), T2 (2,905 cases, 9.0%), and various other classifications in smaller numbers. For the N classification, N0 was predominant (30,802 cases, 95.2%), with smaller counts for N2, N1, and N3. For the M classification, M0 was observed in 31,559 cases (97.6%), with the remainder classified as M1.

Radiation treatment was administered to 1,926 patients (6.0%), whereas 30,422 patients (94.0%) did not receive radiation. The chemotherapy data revealed that 19,480 patients (60.2%) did not undergo chemotherapy, whereas 12,868 patients (39.8%) did.

Patient outcomes revealed that 29,341 patients (90.7%) were alive, whereas 3,007 patients (9.3%) died from urinary bladder cancer.

Model performance

Performance on the test dataset:

The deep neural network model had a sensitivity of 64.55% (95% CI 61.64% to 67.38%). Its specificity was 91.95% (95% CI 90.84% to 92.98%). The positive likelihood ratio was 8.02 (95% CI 6.99 to 9.21), and the negative likelihood ratio was 0.39 (95% CI 0.36 to 0.42). The positive predictive value was 77.43% (95% CI 74.93% to 79.74%), whereas the negative predictive value was 85.85% (95% CI 84.84% to 86.80%). The model's accuracy was 83.75% (95% CI 82.51% to 84.93%). The ROC-AUC was 0.8945 (95% CI 0.8826 to 0.9065). The Brier score was 0.1154.

The logistic regression model had a sensitivity of 64.45% (95% CI 61.54% to 67.29%). The specificity was 91.61% (95% CI 90.47% to 92.65%). The positive likelihood ratio was 7.68 (95% CI 6.71 to 8.79). The negative likelihood ratio was 0.39 (95% CI 0.36 to 0.42). The positive predictive value was 76.65% (95% CI 74.15% to 78.98%). The negative predictive value was 85.77% (95% CI 84.76% to 86.73%). The accuracy was 83.47% (95% CI 82.23% to 84.66%). The ROC-AUC was 0.8862 (95% CI 0.8741 to 0.8987). The Brier score was 0.1196.

The gradient boosting model had a sensitivity of 66.09% (95% CI 63.21% to 68.89%). The specificity was 91.37% (95% CI 90.22% to 92.43%). The positive likelihood ratio was 7.66 (95% CI 6.71 to 8.75). The negative likelihood ratio was 0.37 (95% CI 0.34 to 0.40). The positive predictive value was 76.61% (95% CI 74.15% to 78.90%). The negative predictive value was 86.31% (95% CI 85.29% to 87.26%). The accuracy was 83.80% (95% CI 82.57% to 84.98%). The ROC-AUC was 0.8938 (95% CI 0.8820 to 0.9057). The Brier score was 0.1157.

Performance of the internal validation dataset:

The deep neural network model had a sensitivity of 62.69% (95% CI 59.71% to 65.59%). The specificity reached 91.88% (95% CI 90.77% to 92.91%). The positive likelihood ratio was 7.72 (95% CI 6.73 to 8.86). The negative likelihood ratio was 0.41 (95% CI 0.38 to 0.44). The positive predictive value was 76.10% (95% CI 73.52% to 78.51%). The negative predictive value was 85.66% (95% CI 84.67% to 86.59%). The model's accuracy was 83.36% (95% CI 82.12% to 84.55%). The ROC-AUC was 0.8893 (95% CI 0.8774 to 0.9006). The Brier score was 0.1175.

The logistic regression model had a sensitivity of 61.66% (95% CI 58.67% to 64.58%). The specificity was 92.15% (95% CI 91.05% to 93.16%). The positive likelihood ratio was 7.86 (95% CI 6.83 to 9.04). The negative likelihood ratio was 0.42 (95% CI 0.39 to 0.45). The positive predictive value was 76.42% (95% CI 73.80% to 78.84%). The negative predictive value was 85.36% (95% CI 84.37% to 86.29%). The accuracy was 83.25% (95% CI 82.00% to 84.45%). The ROC-AUC was 0.8827 (95% CI 0.8703 to 0.8944). The Brier score was 0.1205.

The gradient boosting model had a sensitivity of 64.55% (95% CI 61.61% to 67.42%). The specificity was 91.00% (95% CI 89.83% to 92.07%). The positive likelihood ratio was 7.17 (95% CI 6.30 to 8.17). The negative likelihood ratio was 0.39 (95% CI 0.36 to 0.42). The positive predictive value was 74.73% (95% CI 72.20% to 77.11%). The negative predictive value was 86.16% (95% CI 85.16% to 87.11%). The accuracy was 83.28% (95% CI 82.03% to 84.47%). The ROC-AUC was 0.8883 (95% CI 0.8761 to 0.8995). The Brier score was 0.1176.

Deep neural network performance on the external validation dataset:

The deep neural network model had a sensitivity of 69.40% (95% CI 67.72% to 71.05%). Its specificity was 85.32% (95% CI 84.91% to 85.72%). The positive likelihood ratio was 4.73 (95% CI 4.56 to 4.90), and the negative likelihood ratio was 0.36 (95% CI 0.34 to 0.38). The positive predictive value was 32.63% (95% CI 31.84% to 33.44%), whereas the negative predictive value was 96.46% (95% CI 96.27% to 96.64%). The model's accuracy was 83.84% (95% CI 83.43% to 84.24%). The ROC-AUC was 0.8758 (95% CI 0.8698 to 0.8816). The Brier score was 0.1182.

For summary of model performances see Table 1 and Figure 1.

Table 1. Model performance comparison on the test, internal validation, and external validation datasets

Metric	Sensitivity	Specificity	Positive Likelihood Ratio	Negative Likelihood Ratio	Positive Predictive Value	Negative Predictive Value	Accuracy	ROC-AUC	Brier Score
DNN Test	64.55% (95% CI 61.64% to 67.38%)	91.95% (95% CI 90.84% to 92.98%)	8.02 (95% CI 6.99 to 9.21)	0.39 (95% CI 0.36 to 0.42)	77.43% (95% CI 74.93% to 79.74%)	85.85% (95% CI 84.84% to 86.80%)	83.75% (95% CI 82.51% to 84.93%)	0.8945 (95% CI 0.8826 to 0.9065)	0.1154
Logistic Regression Test	64.45% (95% CI 61.54% to 67.29%)	91.61% (95% CI 90.47% to 92.65%)	7.68 (95% CI 6.71 to 8.79)	0.39 (95% CI 0.36 to 0.42)	76.65% (95% CI 74.15% to 78.98%)	85.77% (95% CI 84.76% to 86.73%)	83.47% (95% CI 82.23% to 84.66%)	0.8862 (95% CI 0.8741 to 0.8987)	0.1196
Gradient Boosting Test	66.09% (95% CI 63.21% to 68.89%)	91.37% (95% CI 90.22% to 92.43%)	7.66 (95% CI 6.71 to 8.75)	0.37 (95% CI 0.34 to 0.40)	76.61% (95% CI 74.15% to 78.90%)	86.31% (95% CI 85.29% to 87.26%)	83.80% (95% CI 82.57% to 84.98%)	0.8938 (95% CI 0.8820 to 0.9057)	0.1157
DNN Internal Validation	62.69% (95% CI 59.71% to 65.59%)	91.88% (95% CI 90.77% to 92.91%)	7.72 (95% CI 6.73 to 8.86)	0.41 (95% CI 0.38 to 0.44)	76.10% (95% CI 73.52% to 78.51%)	85.66% (95% CI 84.67% to 86.59%)	83.36% (95% CI 82.12% to 84.55%)	0.8893 (95% CI 0.8774 to 0.9006)	0.1175
Logistic Regression Internal Validation	61.66% (95% CI 58.67% to 64.58%)	92.15% (95% CI 91.05% to 93.16%)	7.86 (95% CI 6.83 to 9.04)	0.42 (95% CI 0.39 to 0.45)	76.42% (95% CI 73.80% to 78.84%)	85.36% (95% CI 84.37% to 86.29%)	83.25% (95% CI 82.00% to 84.45%)	0.8827 (95% CI 0.8703 to 0.8944)	0.1205
Gradient Boosting Internal Validation	64.55% (95% CI 61.61% to 67.42%)	91.00% (95% CI 89.83% to 92.07%)	7.17 (95% CI 6.30 to 8.17)	0.39 (95% CI 0.36 to 0.42)	74.73% (95% CI 72.20% to 77.11%)	86.16% (95% CI 85.16% to 87.11%)	83.28% (95% CI 82.03% to 84.47%)	0.8883 (95% CI 0.8761 to 0.8995)	0.1176
DNN External Validation	69.40% (95% CI 67.72% to 71.05%)	85.32% (95% CI 84.91% to 85.72%)	4.73 (95% CI 4.56 to 4.90)	0.36 (95% CI 0.34 to 0.38)	32.63% (95% CI 31.84% to 33.44%)	96.46% (95% CI 96.27% to 96.64%)	83.84% (95% CI 83.43% to 84.24%)	0.8758 (95% CI 0.8698 to 0.8816)	0.1182

Feature importance

The SHAP summary plots highlight the most influential features in predicting cancer patient mortality. The plots show the average impact of each feature on the model's output, with Ta, Age, Stage I, Size, and Grade IV being the most significant.

For tumor stage, Ta and Stage I had the greatest impact on mortality. Their presence is associated with a lower risk of mortality, whereas their absence increases the risk. Age is also a significant predictor, with higher age values increasing the risk of mortality. Tumor size contributes significantly as well, with larger tumors indicating a greater risk of mortality. Grade IV tumors have a smaller, yet still important, effect on mortality, with patients with anaplastic tumors having a greater chance of death (see Figure 2).

The partial dependence plot revealed that the relationship between age and mortality is not linear but rather quadratic, with greater age significantly increasing the risk of mortality. In contrast, tumor size demonstrated a more linear relationship with mortality. Additionally, the age and size values were positively related to each other (see Figure 3).

Deep neural network architecture

The deep neural network architecture starts with an input layer that expects data with 31 features. This input was processed through three dense layers, each with 100 units, using the leaky ReLU activation function, L2 regularization, and normal initialization, followed by batch normalization. The output layer was a dense layer with a single unit and a sigmoid activation function, which is suitable for binary classification.

The model was compiled via the stochastic gradient descent (SGD) optimizer with a learning rate of 0.001, momentum of 0.9, and Nesterov momentum. The loss function was binary cross-entropy, and the metrics for evaluation were accuracy and AUC.

Early stopping and reducing the learning rate on the plateau were used to optimize the training process. Early training stopped if the validation loss did not improve for 8 epochs, restoring the best weights. The learning rate was reduced when the validation loss plateaued, helping to fine-tune the model further.

This study demonstrated acceptable performance across various models, including a classic statistical model (linear regression), a classic machine learning model (gradient boosting), and a newer deep neural network. Among these models, the deep neural network model slightly outperformed the others on all datasets. The low positive predictive value in the external validation dataset is attributed to the low incidence of deaths due to bladder carcinoma (approximately 9%) in this dataset. In this context, likelihood ratios provide a more suitable measure. The neural network model exhibited a negative likelihood ratio of 0.36, indicating that patients with a negative prediction (i.e., the model predicts that the patient will not die) have a 20–25% lower chance of dying from the disease. Conversely, the positive likelihood ratio of 4.73 increases the pretest probability of death by 25–30% [15]. Additionally, the ROC-AUC of 0.8758 (95% CI: 0.8698 to 0.8816) with a narrow confidence interval demonstrates the model's substantial ability to distinguish between outcomes [16]. Overall, the results of this study suggest that the model could be useful in clinical practice.

Overfitting is a significant problem in machine learning, occurring when a model memorizes the patterns in the training data rather than generalizing from them [17]. Consequently, the model performs well on the initial dataset but fails to predict external data accurately. In this study, external validation with acceptable performance metrics demonstrated that the model is less likely to overfit. Additionally, several techniques have been employed to mitigate overfitting: regularization, early stopping of training, using a limited number of features, and utilizing a simplified neural network [18].

This study has several limitations. First, the sample may not be representative of populations outside the United States. Additionally, nonbinary patients were not included in the study sample. Furthermore, the study predicts outcomes only for postoperative patients, which limits its applicability until after surgery. This may reduce the model's usability for patients who either refuse surgical resection of the tumor or cannot tolerate curative surgery.

The models in this study utilize readily available tumor information to predict outcomes. This approach contrasts with models that require genetic profiling or other specialized data, which may not be accessible, particularly in resource-limited settings.

To improve the usability of our models, we developed a GUI desktop application that leverages a deep neural network model. This application is designed to be user friendly, requiring only basic computer skills to operate. By abstracting away the complexities typically associated with low-level model interactions, which usually necessitate programming knowledge, we have made advanced AI accessible to a broader audience. This represents a significant step toward the widespread adoption and acceptance of artificial intelligence in healthcare.

For future studies, these models can be further validated outside of a U.S. population and in newer, more racially heterogeneous patient groups. With additional information and datasets becoming available, the models trained in this study can be enhanced through transfer learning to leverage the vast amount of data accessible in the era of big data.

Machine learning models can accurately predict mortality in patients with transitional cell carcinoma of the bladder. By utilizing a GUI-based application, the clinical utility of these models in daily practice can be greatly enhanced.

Ethics approval and consent to participate

The access to the dataset is possible only with the permission of corresponding authorities. All patient’s information is deidentified.

Competing Interests

Author Aidin Tarokhian and Mitra Golparian are in a marital relationship. All other authors declare no financial or non-financial competing interests.

Funding

No funding was received for conducting this study.

Author Contributions

All authors have made significant contributions to this work. MG, KS, and AS collected the information of the patients and participated in data retrieval, cleaning and preprocessing. AT and MR contributed to the development and coding of the machine learning models. AT and MG reviewed the dataset and article for potential errors and biases regarding scientific and machine learning principles. Additionally, all authors have been involved in the writing of the article.

Data Availability

The datasets used and/or analyzed during the current study are available online at SEER database. The GUI application for Windows can be shared upon reasonable request from corresponding author.

Code Availability

The underlying code for this study is not publicly available but may be made available to qualified researchers on reasonable request from the corresponding author.

Acknowledgements

We sincerely thank all the people who helped this study.

Manunta A, Vincendeau S, Kiriakou G, Lobel B, Guillé F. Non‐transitional cell bladder carcinomas. BJU international. 2005;95(4):497-502.
van Hoogstraten LM, Vrieling A, van der Heijden AG, Kogevinas M, Richters A, Kiemeney LA. Global trends in the epidemiology of bladder cancer: challenges for public health and clinical practice. Nature reviews Clinical oncology. 2023;20(5):287-304.
Cumberbatch MGK, Jubber I, Black PC, Esperto F, Figueroa JD, Kamat AM, et al. Epidemiology of bladder cancer: a systematic review and contemporary update of risk factors in 2018. European urology. 2018;74(6):784-95.
Simsir A, Sarsik B, Cureklibatir I, Sen S, Gunaydin G, Cal C. Prognostic factors for upper urinary tract urothelial carcinomas: stage, grade, and smoking status. International urology and nephrology. 2011;43:1039-45.
Zang Y, Li X, Cheng Y, Qi F, Yang N. An overview of patients with urothelial bladder cancer over the past two decades: a Surveillance, Epidemiology, and End Results (SEER) study. Annals of Translational Medicine. 2020;8(23).
Lenis AT, Lec PM, Chamie K. Bladder cancer: a review. Jama. 2020;324(19):1980-91.
DeGEORGE KC, Holt HR, Hodges SC. Bladder cancer: diagnosis and treatment. American family physician. 2017;96(8):507-14.
Rajkomar A, Dean J, Kohane I. Machine learning in medicine. New England Journal of Medicine. 2019;380(14):1347-58.
Iniesta R, Stahl D, McGuffin P. Machine learning, statistical learning and the future of biological research in psychiatry. Psychological medicine. 2016;46(12):2455-65.
Cuocolo R, Caruso M, Perillo T, Ugga L, Petretta M. Machine learning in oncology: a clinical appraisal. Cancer letters. 2020;481:55-62.
Bertsimas D, Wiberg H. Machine learning in oncology: methods, applications, and challenges. JCO clinical cancer informatics. 2020;4.
Doll KM, Rademaker A, Sosa JA. Practical guide to surgical data sets: surveillance, epidemiology, and end results (SEER) database. JAMA surgery. 2018;153(6):588-9.
Duggan MA, Anderson WF, Altekruse S, Penberthy L, Sherman ME. The surveillance, epidemiology, and end results (SEER) program and pathology: toward strengthening the critical relationship. The American journal of surgical pathology. 2016;40(12):e94-e102.
Nohara Y, Matsumoto K, Soejima H, Nakashima N. Explanation of machine learning models using shapley additive explanation and application for real data in hospital. Computer Methods and Programs in Biomedicine. 2022;214:106584.
Grimes DA, Schulz KF. Refining clinical diagnosis with likelihood ratios. The Lancet. 2005;365(9469):1500-5.
Çorbacıoğlu ŞK, Aksel G. Receiver operating characteristic curve analysis in diagnostic accuracy studies: A guide to interpreting the area under the curve value. Turkish Journal of Emergency Medicine. 2023;23(4):195.
Kernbach JM, Staartjes VE. Foundations of machine learning-based clinical prediction modeling: Part II—Generalization and overfitting. Machine Learning in Clinical Neuroscience: Foundations and Applications. 2022:15-21.
Gupta GK, Sharma DK. A review of overfitting solutions in smart depression detection models. 2022 9th International conference on computing for sustainable global development (INDIACom): IEEE; 2022. p. 145-51.

Competing interest reported. Author Aidin Tarokhian and Mitra Golparian are in a marital relationship.

Development and External Validation of Machine Learning Models to Predict Postoperative Transitional Cell Carcinoma Mortality: A Retrospective Study

Status:

Version 1

Abstract

Purpose

Methods

Results

Conclusion

Figures

Introduction

Methods and materials

Results

Discussion

Conclusion

Declarations

References

Additional Declarations

Status:

Version 1