Deep-Learning-Based Survival Prediction of Patients with Thyroid Cancer

doi:10.21203/rs.3.rs-3811567/v1

Download PDF

Research Article

Deep-Learning-Based Survival Prediction of Patients with Thyroid Cancer

https://doi.org/10.21203/rs.3.rs-3811567/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

This study acquired data pertaining to individuals diagnosed with Thyroid Cancer (TC) from the Surveillance, Epidemiology, and End Results (SEER) database. Subsequently, a deep learning and neural network model known as DeepSurv was employed to forecast the survival rate of TC patients and assess its efficacy.

Methods

Information on individuals diagnosed with TC from the years 2000 to 2019 was collected from the SEER database. The individuals in question were subsequently allocated into training and testing cohorts through a random selection process, maintaining a ratio of 7:3. The outcomes of the DeepSurv model were compared to those of the Cox proportional-hazards (CoxPH) model in order to estimate the chances of survival for TC patients. The accuracy of the model's predictions was evaluated through the examination of calibration curves, the time-dependent area under the receiver operating characteristic curve (AUC), and the concordance index (C-index).

Results

A total of 25,797 individuals diagnosed with TC were included in this study, with 18,057 comprising the training group and 7,740 forming the testing cohort. The CoxPH model exhibited robust correlations between age, gender, marital status, surgical intervention, radiation therapy, tumor extension, and the survival outcomes of TC patients. Notably, the C-index for the CoxPH model was 0.884, indicating a high level of predictive accuracy. Additionally, the training cohort data were used to create the DeepSurv model, which produced a higher C-index of 0.904. The predictive performance of both models was assessed, and the 3-, 5-, and 8-year AUC values were calculated. Regarding the CoxPH model, the corresponding area under the receiver operating characteristic curve (AUC) values were determined to be 0.835, 0.797, and 0.756, respectively. In comparison, the DeepSurv model achieved higher AUC values of 0.942, 0.918, and 0.906. The DeepSurv model demonstrated superior predictive ability for TC patients, as indicated by both the AUC values and the calibration curve, suggesting higher reliability compared to the CoxPH model.

Conclusion

Using TC patient data from the SEER database for research, we built the DeepSurv model, which performed better than the CoxPH model in estimating the survival time of TC patients.

DeepSurv

Thyroid Cancer

Neural network

Survival prediction

SEER

Thyroid carcinoma represents the predominant neoplastic condition affecting the endocrine system, exhibiting a substantial rise in prevalence within the United States. Notably, it demonstrates the most rapid growth rate among all cancers in the country¹. Prevalence is higher among females, with a female-to-male ratio of 3:1 observed across various geographical regions and demographic cohorts². Furthermore, it ranks as the fifth most prevalent cancer among women. Over the course of the past three decades, the incidence of thyroid cancer has witnessed a staggering surge of 300%, both domestically and globally. It predominantly impacts individuals at a younger age compared to other malignancies, with a median age at diagnosis of 51 years^1,3. Notably, 43% of newly diagnosed cases occur within the age range of 45 to 64 years. Overall, the estimated 5-year survival rate for thyroid cancer stands at an impressive 98.1%, with localized disease exhibiting an even higher rate of 99.9%, while the survival rate drops to 55.5% for cases with distant metastasis². The aforementioned survival rates underscore the significance of timely identification and expeditious intervention. Evaluating the clinical stage of thyroid cancer commonly employs the TNM staging system formulated by the American Joint Committee on Cancer⁴. Although wide local excision performed within 30 days of the initial biopsy has demonstrated mortality reduction in stage I instances, its influence on stage II or III cases is comparatively less pronounced⁵. It is noteworthy that patients with thyroid cancer sharing the same stage can exhibit diverse treatment responses, resulting in divergent therapeutic outcomes and survival rates^6,7.

Advancements in medical research have illuminated the risk factors and prognostic indicators associated with thyroid cancer. Variables such as age, gender, ethnicity, tumor location, and treatment approaches have been correlated with the prognosis of individuals with thyroid cancer^8–10. Novel evaluation methodologies are required to take several contributing factors into account at once in order to more correctly evaluate risk factors and survival rates. In clinical practice, survival analysis methods such as nomograms, logistic regression analyses, and Cox proportional-hazards (CoxPH) models are frequently employed^11–13. However, the linear proportional-hazard models utilised by these techniques could oversimplify the link between predictors and patient prognosis, ignoring temporal fluctuations in the influence of covariates^14–16. Sophisticated algorithms, such as deep learning and neural network models, offer promising avenues in clinical medicine for capturing the multifactorial and nonlinear aspects.

In recent years, a deep neural network model called CoxPH deep neural network (DeepSurv) has been developed to enhance personalized treatment in clinical settings. This innovative model, pioneered by Katzman et al., integrates deep neural networks with the CoxPH model and has exhibited favorable outcomes in clinical applications¹⁷.

Consequently, the objective of this study was to construct a DeepSurv model utilizing patient data from the Surveillance, Epidemiology, and End Results (SEER) database to predict mortality in individuals diagnosed with thyroid cancer. By providing a theoretical foundation for treatment decisions, this model has the potential to significantly improve patient outcomes and facilitate personalized care strategies.

Data filtering criteria

The SEER*Stat software (version 8.4.1) was utilized to screen patients diagnosed with TC^18,19. According to the third version of the International Classification of Diseases for Oncology's primary site codes (C73.9 - Thyroid gland), we particularly chose individuals with primary tumors who had been identified as TC between 2000 and 2019. Patients with nonprimary tumors and inadequate basic data were not included in the analysis. In the end, 25,797 TC patients in all were enrolled in the research. It is significant to highlight that neither ethics committee permission nor signed informed consents were required because, with the exception of patient identification information, all data utilized in the SEER database are available to the general public. The flow chart illustrating the process of patient selection is presented in Fig. 1.

Figure1

Patient information classification criteria

The variables analyzed encompassed factors such as age, gender, ethnicity, marital status, tumor size, tumor spread, TNM stage, summary stage, surgical intervention, radiotherapy treatment, chemotherapy treatment, lymph node dissection following surgery (Reg LN Sur), primary site surgery, sequence of radiotherapy (Rad Seq), and income. The median follow-up duration for this study was 89.57 months, ranging from 1 to 191 months. Race was categorized into White, Black, and Other, while marital status encompassed Married, Single, and Other. Reg LN Sur was divided into three distinct categories: lymph node removal, absence of lymph node inspection, and postoperative lymph node dissection. No radiation, radiotherapy before surgery, radiotherapy during surgery, radiotherapy after surgery, and radiotherapy both before and after surgery were the different categories for Rad Seq.

DeepSurv model design

The DeepSurv model, a feedforward neural network, consists of three layers: the input layer, the hidden layer, and the output layer. It utilizes multiple simulated neurons to process the data. The input layer of the DeepSurv model is composed of the baseline patient data (x). The hidden layer, in turn, incorporates a fully connected nonlinear activation function, dropout regularization, and additional hidden units. The output layer, denoted as h^θ(x)¹⁷ (Fig. 2), provides the estimated risk value. Our model, implemented using the PyTorch deep learning framework, predominantly employed pycox for neural network computations. It leverages clinical characteristics to predict the impact on patient survival and generate a risk value.

We split the TC patients into training and testing groups in order to assess the model. Utilizing the training cohort data, an initial DeepSurv model was constructed using a neural network architecture comprising seven layers. Subsequently, this model was employed to conduct survival analysis on the TC patients within the testing cohort. To assess the discrimination, calibration, and overall efficiency of the DeepSurv models, various measures were employed, including the concordance index (C-index), calibration curve, and receiver operating characteristic (ROC) curve. These measures served as comparative metrics to evaluate the performance of the DeepSurv models.

Figure2

Statistical analysis

Categorical variables were presented as proportions, while continuous variables were summarized using the median and interquartile range. The CoxPH model was developed using R (version 4.2.0), while the DeepSurv model was created using Python (version 3.8.0). Data calculations, model training, and other operations in Python were facilitated by the pandas, matplotlib.pyplot, and NumPy modules²⁰. The deep learning neural network for the DeepSurv model was built using the PyTorch framework and the pycox module. Statistical significance was defined as p < 0.05.

Basic information of patients

The study comprised a total of 25,797 patients diagnosed with TC, who were subsequently divided into two cohorts: a training cohort consisting of 18,057 patients (70%) and a testing cohort consisting of 7,740 patients (30%). 6,820 (26.44%) men and 18,977 (73.56%) women made up the patient population. The average age of the patients was 50 years, and the majority, 20,879 (80.94%), were of white ethnicity. Regarding the disease stage, most patients were classified as T1 (13,222, 51.25%), N0 (18,773, 72.77%), and M0 (24,563, 95.22%). The median duration of follow-up was 89.57 months, with a range spanning from 1 to 191 months. There were 3,873 (3.38%) deaths attributed to TC. Notably, the survival curves and essential clinical data exhibited no significant differences between the two cohorts. For a comprehensive overview of the patient groups' fundamental characteristics, please refer to Table 1, while Fig. 3 presents the Kaplan-Meier analysis curve.

Table 1

Baseline Characteristics of Thyroid Cancer
Variable	Total N(%)	Train cohort N(%)	Test cohort N(%)
total	25797(100%)	18057(100%)	7740(100%)
Age(years old)
mean ± sd	50 ± 16.31	50 ± 16.31	50 ± 16.33
CS_tumor_size(mm)
mean ± sd	52 ± 169.3	52 ± 169.13	52 ± 169.71
CS_extension
mean ± sd	267 ± 186.25	265 ± 185.24	271 ± 188.53
Surg_Prim_Site
mean ± sd	44 ± 12.89	44 ± 12.8	44 ± 13.1
Sex
male	6820(26.44%)	4770(26.42%)	2050(26.49%)
female	18977(73.56%)	13287(73.58%)	5690(73.51%)
Race
white	20879(80.94%)	14574(80.71%)	6305(81.46%)
black	1686(6.54%)	1205(6.67%)	481(6.21%)
other	3232(12.53%)	2278(12.62%)	954(12.33%)
Histology type
follicular	1811(7.02%)	1262(6.99%)	549(7.09%)
papillary	21581(83.66%)	15145(83.87%)	6436(83.15%)
anaplastic	902(3.5%)	611(3.38%)	291(3.76%)
medullary	214(0.83%)	149(0.83%)	65(0.84%)
other	1289(5%)	890(4.93%)	399(5.16%)
Marital
single	5582(21.64%)	3945(21.85%)	1637(21.15%)
married	16589(64.31%)	11592(64.2%)	4997(64.56%)
DSW	3626(14.06%)	2511(13.91%)	1115(14.41%)
Grade
I	17037(66.04%)	11364(62.93%)	5673(73.29%)
II	3829(14.84%)	2670(14.79%)	1159(14.97%)
III	1425(5.52%)	994(5.5%)	431(5.57%)
IV	1506(5.84%)	1029(5.7%)	477(6.16%)
AJCC
I	16828(65.23%)	11814(65.43%)	5014(64.78%)
II	1963(7.61%)	1366(7.56%)	597(7.71%)
III	3381(13.11%)	2377(13.16%)	1004(12.97%)
IV	3625(14.05%)	2500(13.85%)	1125(14.53%)
T
T0	13(0.05%)	11(0.06%)	2(0.03%)
T1	13222(51.25%)	9236(51.15%)	3986(51.5%)
T2	4420(17.13%)	3114(17.25%)	1306(16.87%)
T3	5350(20.74%)	3780(20.93%)	1570(20.28%)
T4	2629(10.19%)	1807(10.01%)	822(10.62%)
TX	163(0.63%)	109(0.6%)	54(0.7%)
N
N0	18773(72.77%)	13178(72.98%)	5595(72.29%)
N1	6723(26.06%)	4675(25.89%)	2048(26.46%)
NX	301(1.17%)	204(1.13%)	97(1.25%)
M
M0	24563(95.22%)	17207(95.29%)	7356(95.04%)
M1	1172(4.54%)	812(4.5%)	360(4.65%)
M2	62(0.24%)	38(0.21%)	24(0.31%)
Summary_Stage
localized	16291(63.15%)	11435(63.33%)	4856(62.74%)
regional	7517(29.14%)	5273(29.2%)	2244(28.99%)
distant	1989(7.71%)	1349(7.47%)	640(8.27%)
Surgery
yes	24887(96.47%)	17443(96.6%)	7444(96.18%)
no	910(3.53%)	614(3.4%)	296(3.82%)
Radiation
yes	13331(51.68%)	9302(51.51%)	4029(52.05%)
no	12466(48.32%)	8755(48.49%)	3711(47.95%)
Chemotherapy
yes	756(2.93%)	519(2.87%)	237(3.06%)
no	25041(97.07%)	17538(97.13%)	7503(96.94%)
Reg_LN_Sur
yes	13649(52.91%)	9574(53.02%)	4075(52.65%)
no	12148(47.09%)	8483(46.98%)	3665(47.35%)
Rad_Seq
before	81(0.31%)	52(0.29%)	29(0.37%)
after	12528(48.56%)	8755(48.49%)	3773(48.75%)
intraoperative	4(0.02%)	1(0.01%)	3(0.04%)
both	58(0.22%)	39(0.22%)	19(0.25%)
no	13126(50.88%)	9210(51.01%)	3916(50.59%)
Income
low	5535(21.46%)	3877(21.47%)	1658(21.42%)
mediate	11542(44.74%)	8007(44.34%)	3535(45.67%)
high	8720(33.8%)	6173(34.19%)	2547(32.91%)
a CS tumor size: Information on tumor size. Available for after 2004 year. Earlier cases may be converted and new codes added which weren’t available for use prior to the current version of CS.
b CS extension: Information on extension of the tumor. Available for after 2004 year. Earlier cases may be converted and new codes added which weren’t available for use prior to the current version of CS.
c Surg_Prim_Site: Surgery of Primary Site describes a surgical procedure that removes and/or destroys tissue of the primary site performed as part of the initial work-up or first course of therapy.
d Reg_LN_Sur: Scope of Regional Lymph Node Surgery describes the procedure of removal,biopsy, or aspiration of regional lymph nodes performed during the initial work-up or first course of therapy at all facilities.
e Rad_Seq: This field records the order in which surgery and radiation therapies were administered for those patients who had both surgery and radiation.

Table 1

Figure3

Variable screening and DeepSurv model training

Employing the CoxPH model to analyze the multivariate factors within the training cohort unveiled the influential risk factors contributing to patient mortality, including age, sex, marital status, surgery, radiotherapy, and tumor extension (Table 2). Notably, the CoxPH model exhibited a C-index of 0.884. In parallel, we developed the DeepSurv model using the training cohort, achieving a commendable C-index of 0.904. The training loss-function diagram, depicted in Fig. 4, visually illustrates the model's performance. Upon comparison, it is evident that the DeepSurv model outperformed the CoxPH model in terms of effectiveness.

Table 2

Survival predictors in Cox PH model
Variables	β	HR	95%CI	P
Age	0.06825	1.071	1.068–1.073	< .0001***
Sex female	-0.44863	0.639	0.597–0.683	< .0001***
Marital married	-0.29979	0.741	0.674–0.815	< .0001***
Maritaldsw	-0.05263	0.949	0.85–1.059	0.3464
Surgeryno	1.68626	5.399	4.838–6.026	< .0001***
Radiationno	-0.00365	0.996	0.934–1.063	0.0124**
CS_extension	0.00244	1.002	1.002–1.003	< .0001***

Table 2

Figure4

Comparison of the DeepSurv model and CoxPH model in the testing cohort

To check the precision of the CoxPH model's (Fig. 5) and the DeepSurv model's (Fig. 6) capacity to estimate survival probability, we first built calibration curves for patients diagnosed with TC at 3, 5, and 8 years. By plotting ROC curves for TC patients at 3, 5, and 8 years into the future, the discrimination between the two models can be assessed. Specifically, the time-dependent area under the ROC curve (AUC) value can be calculated to evaluate their performance (Fig. 7). The outcomes indicated that the DeepSurv model exhibited superior performance compared to the Cox model, as evidenced by its higher AUCs at 3, 5, and 8 years (0.942, 0.918, and 0.906, respectively) relative to the Cox model (0.835, 0.797, and 0.756, respectively). These findings unequivocally demonstrate that the DeepSurv model surpasses the CoxPH model in terms of its prognostic accuracy and calibration capabilities for predicting the survival prognosis of patients with TC.

Figure5

Figure6

Figure7

Thyroid cancer commands global attention as a consequential public health challenge, characterized by disparities in morbidity and mortality rates across diverse geographical regions and populations. Global estimates underscore the gravity of the situation, with an annual incidence exceeding 232,100 new cases, representing approximately 1.7% of the global cancer landscape²¹. Regrettably, TC-related mortality claims the lives of approximately 55,500 individuals each year, accounting for approximately 0.7% of cancer-related deaths²¹. The observed heterogeneity in the morbidity and mortality patterns of thyroid cancer finds its roots in multifaceted factors, including differential access to primary healthcare, variances in early detection strategies, and disparities in treatment modalities^22,23. Furthermore, the advent of metastasis precipitates a stark decline in the prognosis of thyroid cancer, heralding a considerable reduction in survival rates^21,24.

This lamentable reality underscores the pressing need for innovative therapeutic strategies and comprehensive management approaches. An additional dimension that warrants consideration in the realm of thyroid cancer is the evident heterogeneity in mortality rates based on factors such as country of residence, race, and socioeconomic status^21,22,25. Consequently, beyond the realm of treatment options, it becomes incumbent upon healthcare practitioners and researchers to embrace a holistic approach that acknowledges the influence of environmental factors and the psychological well-being of patients. Only through such a comprehensive paradigm can we envisage the enhancement of survival rates and amelioration in the overall quality of life for individuals confronted with the challenges of thyroid cancer²⁶.

An array of survival prediction models has been formulated to elevate the precision of prognosticating patient survival times. Although the Cox proportional hazards (CoxPH) model is widely employed, its efficacy is hampered by the linearity of its constituents. Conversely, the DeepSurv approach has garnered commendation and found application across diverse subspecialties within clinical medicine¹⁷. Multiple investigations have demonstrated the superior performance of the DeepSurv model over conventional linear prediction models in prognosticating survival outcomes¹⁷. Notably, empirical evidence substantiates the heightened accuracy of the DeepSurv model compared to the CoxPH model in diverse malignancies, encompassing lung cancer, colon adenocarcinoma, and patients within Coronary Care Units^27–31.

In this investigation, a cohort of patients with TC was meticulously divided into distinct subsets to facilitate comprehensive analysis and model establishment. Notably, a noteworthy proportion of 70% was devoted to the training cohort, meticulously harnessed to execute the multivariate analysis of the esteemed CoxPH model and subsequently establish the esteemed DeepSurv model. The remaining 30% was aptly assigned to the test cohort, effectively serving as a crucible to evaluate and validate the predictive capacities of both models.

Within the purview of the CoxPH model, an assortment of factors including age, sex, marital status, surgical interventions, radiotherapy, and tumor extension were aptly delineated as substantive risk elements impinging on the domain of TC, as distinctly expounded upon in Table 2. Furthermore, the CoxPH model rendered an impressive C-index of 0.884, attesting to its commendable predictive precision.

Concomitantly, the nascent minted DeepSurv model, embodying an intricate neural network comprising seven discerning layers, remarkably attained a C-index of 0.904. Notably, an intriguing disparity emerged between the calibration curves of the DeepSurv and CoxPH models. The former, boasting a more evenly distributed profile harmoniously aligned with the leading-diagonal line, exuded an unmistakable aura of superiority. Evident in the realm of the AUC curve, the DeepSurv model exhibited a remarkable smoothness that transcended its CoxPH counterpart, reiterating its unassailable prowess when discerning and scrutinizing the influential elements dictating 3-, 5-, and 8-year mortality and survival-time prognostications for patients afflicted with TC.

Intriguingly, the superior AUC curve, elevated above that of the CoxPH model, serves as a poignant testament to the exalted predictive and discriminative aptitude of the DeepSurv model. Its compelling superiority stems from the adroit employment of multilevel neural networks, adroitly confronting the intricacies encompassing expansive sample sizes, intricate multifactorial variables, and the inherent nonlinearity innate to the realm of prognostic predictions. Consequently, it is irrefutably apparent that the DeepSurv model holds sway as a formidable contender, imbued with inherent advantages when discerning the precise survival prognosis for patients beset by TC, surpassing the capabilities exhibited by its counterparts.

This study encountered several limitations that warrant acknowledgment. Firstly, the absence of crucial prognostic factors, such as intricate surgical techniques, tailored radiotherapy protocols, specific chemotherapy regimens, pharmacological interventions, and related details, in the SEER database constrained the comprehensiveness of our findings in relation to patients with TC. Secondly, the absence of external validation restricted the generalizability of our research findings, as the dataset exclusively encompassed information from select states within the United States. Future endeavors will strive to enhance the DeepSurv model by incorporating more expansive and diverse datasets from a broader geographical scope. Thirdly, the inherent opacity of the hidden layer in the DeepSurv model, acting as a computational "black box," posed challenges in elucidating the precise mechanisms underpinning its predictive capacity and the decision-making process it employs. In our forthcoming investigation, we endeavor to tackle these aforementioned limitations through meticulous exploration and elucidation.

This study marked a pioneering endeavor in constructing a dedicated DeepSurv prediction model tailored to the precise estimation of survival time among patients diagnosed with TC. Notably, the DeepSurv model exhibited substantial superiority over the conventional CoxPH model in prognosticating the determinants that impact patient survival. Thus, the DeepSurv model emerges as an innovative analytical instrument, affording valuable insights into outcome prediction and treatment recommendations for individuals grappling with TC.

Conflicts of Interest

The authors declare no conflict of interest. The funders had no role in the design of the study.

Ethics approval and consent to participate

Not applicable

Data availability statement

Publicly available datasets were analyzed in this study. This data can be found at: https://seer.cancer.gov.

Acknowledgments

We thank all SEER database staff and scientists. We are also very grateful to the Guangdong Provincial Key Laboratory of Traditional Chinese Medicine Informatization for their support.

Funding

This research was funded by Key Scientific Problems and Medical Technical Problems Research Project of China Medical Education Association (grant number 2022KTZ009) and Guangdong Provincial Key Laboratory of Traditional Chinese Medicine Informatization (grant number 2021B1212040007).

Miranda-Filho A, Lortet-Tieulent J, Bray F, et al. Thyroid cancer incidence trends by histology in 25 countries: a population-based study. The lancet Diabetes & endocrinology. 2021;9(4):225–34.
Siegel RL, Miller KD, Fuchs HE, Jemal A, Cancer statistics. 2022. CA: a cancer journal for clinicians. 2022;72(1):7–33.
Roman BR, Morris LG, Davies L. The thyroid cancer epidemic, 2017 perspective. Curr Opin Endocrinol Diabetes Obes. 2017;24(5):332–6.
Lincango-Naranjo E, Solis-Pazmino P, El Kawkgi O, et al. Triggers of thyroid cancer diagnosis: a systematic review and meta-analysis. Endocrine. 2021;72(3):644–59.
Araque KA, Gubbi S, Klubo-Gwiezdzinska J. Updates on the Management of Thyroid Cancer. Hormone and metabolic research = Hormon- und Stoffwechselforschung = Hormones et metabolisme. 2020;52(8):562–77.
Filetti S, Durante C, Hartl D, et al. Thyroid cancer: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up†. Annals of oncology: official journal of the European Society for Medical Oncology. 2019;30(12):1856–83.
Nabhan F, Dedhia PH, Ringel MD. Thyroid cancer, recent advances in diagnosis and therapy. Int J Cancer. 2021;149(5):984–92.
Miller KD, Fidler-Benaoudia M, Keegan TH, Hipp HS, Jemal A, Siegel RL. Cancer statistics for adolescents and young adults, 2020. Cancer J Clin. 2020;70(6):443–59.
Shukla N, Osazuwa-Peters N, Megwalu UC. Association Between Age and Nodal Metastasis in Papillary Thyroid Carcinoma. Otolaryngology–head and neck surgery: official journal of American Academy of Otolaryngology-Head and Neck Surgery. 2021;165(1):43–9.
Ma GMY, Makary MS, Shujaat TM, Prevedello LM, Erdal SBS, Nguyen XV. Neck CT imaging and correlation with thyroid cancer incidence across age, gender and race. Clin Endocrinol. 2021;94(5):872–9.
Christensen E. Multivariate survival analysis using Cox's regression model. Hepatology (Baltimore MD). 1987;7(6):1346–58.
Wu M, Ou-Yang DJ, Wei B, et al. A Prognostic Model of Differentiated Thyroid Cancer Based on Up-Regulated Glycolysis-Related Genes. Front Endocrinol. 2022;13:775278.
Ruiz EML, Niu T, Zerfaoui M, et al. A novel gene panel for prediction of lymph-node metastasis and recurrence in patients with thyroid cancer. Surgery. 2020;167(1):73–9.
Asano J, Hirakawa A, Hamada C. Assessing the prediction accuracy of cure in the Cox proportional hazards cure model: an application to breast cancer data. Pharm Stat. 2014;13(6):357–63.
Camp NJ, Madsen MJ, Herranz J, et al. Re-interpretation of PAM50 gene expression as quantitative tumor dimensions shows utility for clinical trials: application to prognosis and response to paclitaxel in breast cancer. Breast Cancer Res Treat. 2019;175(1):129–39.
Goerdten J, Carrière I, Muniz-Terrera G. Comparison of Cox proportional hazards regression and generalized Cox regression models applied in dementia risk prediction. Alzheimer's & dementia (New York N Y). 2020;6(1):e12041.
Katzman JL, Shaham U, Cloninger A, Bates J, Jiang T, Kluger Y. DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Med Res Methodol. 2018;18(1):24.
Wu WT, Li YJ, Feng AZ, et al. Data mining in clinical big data: the frequently used databases, steps, and methodological models. Military Med Res. 2021;8(1):44.
Yang J, Li Y, Liu Q, et al. Brief introduction of medical database and data mining technology in big data era. J evidence-based Med. 2020;13(1):57–69.
Munir K, Elahi H, Ayub A, Frezza F, Rizzi A. Cancer Diagnosis Using Deep Learning: A Bibliographic Review. Cancers 2019;11(9).
(n.d.) WCRFI. Thyroid cancer statistics. i>https://wwwwcrforg/cancer-trends/worldwide-cancer-data/. 2022.
Grønlund MP, Jensen JS, Hahn CH, Grønhøj C, Buchwald CV. Risk Factors for Recurrence of Follicular Thyroid Cancer: A Systematic Review. Thyroid: official journal of the American Thyroid Association. 2021;31(10):1523–30.
Suteau V, Munier M, Briet C, Rodien P. Sex Bias in Differentiated Thyroid Cancer. Int J Mol Sci 2021;22(23).
Carling T, Udelsman R. Thyroid cancer. Annu Rev Med. 2014;65:125–37.
Miller KD, Nogueira L, Devasia T, et al. Cancer treatment and survivorship statistics, 2022. Cancer J Clin. 2022;72(5):409–36.
Laha D, Nilubol N, Boufraqech M. New Therapies for Advanced Thyroid Cancer. Front Endocrinol. 2020;11:82.
Krittanawong C, Zhang H, Wang Z, Aydar M, Kitai T. Artificial Intelligence in Precision Cardiovascular Medicine. J Am Coll Cardiol. 2017;69(21):2657–64.
Ching T, Zhu X, Garmire LX. Cox-nnet: An artificial neural network method for prognosis prediction of high-throughput omics data. PLoS Comput Biol. 2018;14(4):e1006076.
She Y, Jin Z, Wu J, et al. Development and Validation of a Deep Learning Model for Non-Small Cell Lung Cancer Survival. JAMA Netw open. 2020;3(6):e205842.
Yang R, Huang T, Wang Z et al. Deep-Learning-Based Survival Prediction of Patients in Coronary Care Units. Computational and mathematical methods in medicine. 2021;2021:5745304.
Yu H, Huang T, Feng B, Lyu J. Deep-learning model for predicting the survival of rectal adenocarcinoma patients based on a surveillance, epidemiology, and end results analysis. BMC Cancer. 2022;22(1):210.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Deep-Learning-Based Survival Prediction of Patients with Thyroid Cancer

Status:

Version 1

Abstract

Background

Methods

Results

Conclusion

Figures

Background

Materials and Methods

Data filtering criteria

Patient information classification criteria

DeepSurv model design

Statistical analysis

Results

Basic information of patients

Variable screening and DeepSurv model training

Comparison of the DeepSurv model and CoxPH model in the testing cohort

Discussion

Conclusions

Declarations

References

Additional Declarations

Status:

Version 1