Machine Learning-Based Prediction of Gleason Grade Group upgrading in Patients with Localized Prostate Cancer Awaiting Surgery

doi:10.21203/rs.3.rs-4970391/v1

Download PDF

Article

Machine Learning-Based Prediction of Gleason Grade Group upgrading in Patients with Localized Prostate Cancer Awaiting Surgery

https://doi.org/10.21203/rs.3.rs-4970391/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background: Despite the improved precision of the MRI fusion prostate biopsy, discrepancies persist between the Gleason grade group (GG) biopsy and the pathological Gleason GG. Our study employs machine learning to predict the upgrading of the Gleason GG, aiding treatment decisions.

Material & Methods: Since 2009, we retrospectively reviewed localized prostate cancer patients who underwent prostatectomy, considering seven potential factors contributing to the upgrading: age, prostate specific antigen (PSA) level, PSA density, biopsy GG, Prostate Imaging-Reporting and Data Systems, percent positive cores and surgical waiting time. Pearson'scorrelation and principal component analysis(PCA) were used to explore the data. Various machine learning models were employed for comparison.

Results: Of 418 patients, neither the Pearson correlation nor the PCA revealed strong correlations with GG upgrading. Logistic regression (LR) achieved the best F1 score, though all models had F1 scores below 0.5, indicating prediction challenges. LR and Neural Network analysis identified biopsy GG, age, and percent positive cores as significant predictors.

Conclusions: No specific features strongly correlated with GG upgrading. Despite high accuracy, intelligent concepts struggled to predict upgrades effectively. Physician expertise and patient characteristics remain crucial for management decisions. We agree that machine learning has great potential to improve prediction in the future.

Biological sciences/Cancer/Urological cancer/Prostate cancer

Biological sciences/Computational biology and bioinformatics/Machine learning

Gleason score discordance

upgrading

prostate cancer

machine learning

Prostate cancer (PCa) is the fourth most common cancer among men in Thailand(1–3). The wide range of treatment strategies was based on the risk assessment of individual including the prostate-specific antigen (PSA) level, and Gleason score or The International Society of Urological Pathology (ISUP) grade group (GG) from biopsy(4). Despite the increased precision of the MRI fusion prostate biopsy(5), the discordance persists between the biopsy GG and the pathological GG, which can significantly lead to an increased risk of positive surgical margins, biochemical recurrence, and overall survival prognosis for PCa(6, 7). Furthermore, underestimating biopsy GG can mislead clinical decision making to the patient undergoing active surveillance choice. Therefore, it is crucial to identify the primary factor responsible for these score upgrades.

Numerous studies have explored the potential factors contributing to this discordance. Patient characteristics, pathological factors, and biopsy techniques have undergone extensive assessment over the years to predict GG upgrades. Most research indicates that older age, PSA, PSA density (PSAD), and percentage of positive cores are related to upgrading (8–11). Additionally, Prostate Imaging-Reporting and Data Systems (PI-RADS) was also mentioned as a predictor in the study of Song et al(12). Although previous research has provided valuable information on various contributing factors, there is a significant research gap, specifically the influence of the waiting period for radical prostatectomy (RP) on GG upgrades.

Considering the substantial number of patients waiting for surgery at Siriraj Hospital, the question arises of whether combining the time-related feature will help predict the upgrading. Existing studies by Filipou et al.(13) and Weiner B. et al.(14) have shown no substantial differences in pathology outcomes between immediate (within 6 months) and delayed (beyond 6 months) RP; however, RP beyond 1 year exhibited a more pronounced impact. Based on these findings, we have incorporated traditional predictors and time-related characteristics, including age, PSA, PSAD, PI-RADS, percent positive core, GG at biopsy, and surgical waiting time, into machine learning (ML) analysis to improve the prediction of GG upgrading.

ML, which consists of mathematics, statistics, and data science, has gained significant popularity in the field of medicine, particularly for helping diagnose and predicting clinical outcomes. Although various statistical analyzes often concentrate on relationships between variables, ML prioritizes precise predictions through algorithmic training with data(15–17). Therefore, our study used ML to predict the probability of GG upgrades based on possible causes, including clinical, pathological parameters, and waiting period for surgery, in order to assist physicians in treatment decision making.

Patient selection, Data preparation, and Study parameters

With the approval of the Ethics Committee of the Siriraj Institutional Review Board, a retrospective analysis of the dataset at Siriraj Hospital was conducted in accordance with relevant regulations and guidelines. Due to the retrospective nature of the study, the Ethics Committee of the Siriraj Institutional Review Board waived the requirement for informed consent.

We conducted a retrospective review of medical records, since January 2019, at Siriraj Hospital for patients diagnosed with localized PCa, confirmed by any biopsy methods, and subsequently underwent RP. Initially, the study included 558 patients. However, 133 were excluded due to incomplete data and 7 more were excluded due to a waiting surgery period that exceeded 1,000 days, considered the outlier, involving patient decision making problems. After these exclusions, the study consisted of a total of 418 patients.

Patient data were categorized into two classes for ML-based analysis: Class 0 represented patients with a non-upgrading GG, while Class 1 included those with an upgrading GG, indicating that pathological GG is greater than biopsy GG. The sample size for the ML analysis was determined based on 7 variables using the equation for binary outcomes, proposed by Riley et el.(18) to ensure accurate predictions and prevent overfitting. The study parameters included age, PSA, PSAD, PI-RADS ranging from 1 to 5, percent positive core, GG at biopsy ranging from 1 to 4, and surgical waiting time.

Machine Learning Analysis

After defining the dataset as binary classification tasks, where 1 implies GG upgrading, and 0 implies otherwise, we conducted our data exploratory analysis using Orange Data Mining(19) for its user-friendly interface, simplicity in deployment, and its ability to generate visually appealing representations of the data. To examine the relationship between features, Pearson's correlation was employed. Correlation values range from − 1 to 1, where − 1 indicates a strong negative correlation, 1 indicates a strong positive correlation and 0 indicates no correlation. The degrees of correlation, described as negligible, weak, moderate and strong, were defined as mentioned in Schober et al. (20). Additionally, for data visualization and dimensionality reduction, we utilized Principal Component Analysis (PCA). The cumulative percentage of explained variance from the chosen components exceeding 80% can empirically represent the dataset in the data exploration task, while aiming to select as few components as possible(21). In addition, we also used the Explained add-on library on Orange Data Mining to interpret the factors influencing the predictions after model evaluation.

In model training and experimentation, we employed a variety of ML models, including Naïve Bayes, XGBoost, Random Forest (RF), Support Vector Machines (SVM), Logistic Regression (LR), and Neural Network (NN) for comparison(22–24). To assess the model's performance, we opted for stratified 10-fold cross-validation(24). Parameter tuning was carried out within each fold using a 3-fold cross-validation. Addressing the slight data imbalance with the proportion of the minority class at 20.81%, we used over-sampling and under-sampling techniques within each fold of the cross-validation process(25, 26). This allowed us to compare outcomes across different scenarios: oversampling, undersampling, and no sampling. Evaluation metrics included accuracy, precision, recall, and F1 score. Regarding baseline accuracy, since the majority of our dataset was classified as Class 0, we would have a baseline model accuracy of 82.1% if we labeled everyone as Class 0. All these processes were managed using Pandas(27) for data handling and Scikit-learn(28) for ML methods and algorithms.

Statistical Analysis

Demographic data were presented as mean and standard deviation for age, while median and interquartile ranges were used for PSA level, PSAD, surgical waiting time and percent positive cores. To compare normally distributed continuous data, independent t tests were applied, while nonnormally distributed data were assessed using the Mann-Whitney U test. Qualitative data were analyzed with Chi-square and Fisher's exact tests. Statistical analyzes were also performed using Stata/MP version 16.1 for Mac (StataCorp. 2020, College Station, TX, USA). Statistical significance was defined as a p-value less than 0.05.

Table 1

Overall characteristics of 418 patients with PCa who underwent RP.
Parameters	Total (N = 418)	Nonupgrades (Class 0) (n = 331)	Upgrades (Class 1) (n = 87)	p-value
Age^a, years		68.8 (6)	68.6 (7.3)	0.83
PSA level^b, ng/ml		8.7 (6.2, 13.2)	9.3 (6.9, 14.8)	0.16
PSAD^b, ng/ml²		0.3 (0.2, 0.4)	0.3 (0.2, 0.4)	0.8
PI-RADS^c				< 0.01
3	25 (6)	12 (3.6)	13 (14.9)
4	258 (61.7)	207 (62.5)	51 (58.6)
5	135 (32.3)	112 (33.8)	23 (26.4)
Surgical waiting time^b, weeks		11.3 (7.1, 18.1)	13.1 (8.1, 17.9)	0.27
Percent positive core^b, %		36.8 (21.4, 54.6)	20 (9.1,37.5)	< 0.01
Biopsy GG^b				< 0.01
1 (GS \(\:\le\:\) 6)	61 (14.6)	20 (6)	41 (47.1)
2 (GS 7 [3 + 4])	149 (35.7)	124 (37.5)	25 (28.7)
3 (GS 7 [4 + 3])	74 (17.7)	63 (19)	11 (12.6)
4 (GS 8)	52 (12.4)	42 (12.7)	10 (11.5)
5 (GS 9)	82 (19.6)	82 (24.8)	0
^a Data are presented as mean ± standard deviation and p-value is calculated by independent T-test. ^b Data are presented as median (interquartile range) and p-value is calculated by Mann-Whitney U test. ^c Data are presented as n (%) and p-value is calculated by Chi-square or Fisher’s exact test.

Patient Characteristics

Among the 418 patients diagnosed with clinically localized PCa who underwent RP, 331 in the nonupgrade group were labeled Class 0, and the remaining 87 in the upgrade group were labeled Class 1. Table 1 provides an overview of the characteristics of these patients. The mean age of 331 patients in Class 0 and patients in Class 1 were 68.75 (±5.95) and 68.59 (±7.26) years, respectively. The median level of preoperative PSA was 8.74 (6.2, 13.15) ng/ml in Class 0 and 9.3 (6.9, 14.8) ng/ml in Class 1. The median PSAD was approximately 0.25 ng/ml² for both classes. Within the PI-RADS score, the majority in both classes fell into PI-RADS 4 (207 vs. 51), followed by PI-RADS 5 (113 vs. 23), and finally PI-RADS 3 (12 vs. 13) with a p-value less than 0.01. The median surgical waiting time was 11.29 (7.14, 18.14) weeks for Class 0 and 13.14 (8.14, 17.86) weeks for Class 1. The median percent of positive cores were 36.84 (21.43, 54.55) and 20 (9.09, 37.5) with a p-value < 0.01. For Class 0, most patients were confirmed with ISUP GG 2 (126, 37.72%), while for Class 1, ISUP GG 1 was more prevalent (41, 47.13%) at biopsy, with a p-value < 0.01.

Table 2

Pearson correlation results indicating the association between features in the dataset
Variables	Age	PSA	PSAD	PI-RADS	Surgical waiting time	Percent positive core	Biopsy GG	GG upgrading
Age	1
PSA	-0.0851 0.0822	1
PSAD	-0.1350 0.0057	0.8255 < 0.0001	1
PI-RADS	-0.1060 0.0302	-0.1936 0.0001	-0.2411 < 0.0001	1
Surgical waiting time	0.0224 0.6485	-0.0058 0.9053	0.0008 0.9870	-0.0515 0.2933	1
Percent positive core	0.0343 0.4844	0.2765 < 0.0001	0.3863 < 0.0001	-0.3976 < 0.0001	-0.0710 0.1471	1
Biopsy GG	-0.1542 0.0016	-0.1817 0.0002	-0.2582 < 0.0001	0.3961 < 0.0001	0.0230 0.6385	-0.4795 < 0.0001	1
GG upgrading	0.0104 0.8316	-0.0141 0.7744	0.0311 0.5258	-0.1357 0.0055	0.0018 0.9705	0.2393 < 0.0001	-0.3725 < 0.0001	1
P-values are reported beneath the corresponding Pearson correlation coefficients. A p-value less than 0.05 is considered statistically significant.

Data Exploration

As demonstrated in Table 2, Pearson's correlation analysis focused on the target variable, GG upgrade, showed that biopsy GG had the highest correlation score with a weak negative correlation of 0.3725 (p-value < 0.0001), followed by the percentage of positive core with a weak positive association of 0.2393 (p-value < 0.0001), and then PI-RADS with a mild negative correlation score of 0.1357 (p-value 0.0055). Although the analysis did not reveal specific features strongly associated with the upgrade of GG, we proceeded with further exploration using PCA.

Figure 1A presents the scree plot, revealing that principal component (PC) 1 captures the highest variance at 0.35%, followed by PC2 at approximately 0.12%, and the remaining components show similar values. Achieving more than 80% explained variance requires the inclusion of up to 7 PCs. In Fig. 1B, the 2D visualization using only two PCs on the x and y axes demonstrates difficulty in distinguishing between classes (Class 0 and 1).

Model Performances Across Sampling Strategies

The performance of different models under various strategies to address data imbalance is shown in Table 3: Oversampling results, undersampling results, and no sampling results. When considering the oversampling method, the training accuracy varies from 0.621 to 1, with XGBoost and RF achieving a perfect score of 1, followed by NN with 0.8279. The test precision ranges from 0.6047 to 0.8037, where RF leads with the highest accuracy of 0.8037, followed by XGBoost (0.7698) and LR (0.7171). The precision varies from 0.2972 to 0.5260, with RF achieving the highest precision at 0.5260, followed by XGBoost (0.4177) and LR (0.3891). Examining recall, Naïve Bayes leads with 0.8014, followed by LR (0.6028) and SVM (0.5235). The F1 score ranges from 0.3443 to 0.4601, with LR the highest at 0.4601, followed by RF (0.4438) and Naïve Bayes (0.4327).

Following the undersampling method, the training precision ranges from 0.6705 to 0.9529, with RF achieving the highest accuracy (0.9529), followed by XGBoost (0.8364) and SVM (0.7768). The test accuracy ranges from 0.6234 to 0.7169, with XGBoost leading at 0.7169, followed by RF (0.7027) and SVM (0.6714). The precision ranges from 0.152 to 0.3734, with XGBoost achieving the highest at 0.3734, followed by RF (0.3727) and SVM (0.3526). Naïve Bayes led in recall with 0.7583, followed by SVM (0.7112) and LR (0.6946). SVM achieved the highest F1 score of 0.4680, followed by RF (0.4592) and Naïve Bayes (0.4509).

Regarding the No-sampling method, the training accuracy ranges from 0.7106 to 0.9060, with RF achieving the highest accuracy (0.9060), followed by XGBoost (0.8532) and LR (0.8335). The accuracy of the test ranges from 0.6466 to 0.8277, with LR leading at 0.8277, followed by RF (0.8229) and XGBoost (0.8226). The precision ranges from 0.3889 to 0.6533, with LR leading at 0.6533, followed by Naïve Bayes (0.6353) and RF (0.6167). In terms of recall, NN led with 0.6306, followed by SVM (0.5639) and Naïve Bayes (0.3816). NN achieved the highest F1 score of 0.4923, followed by LR (0.4278) and SVM (0.4041).

In addition, we examine the features that influence the model output. We selected two models: LR, which achieves the highest accuracy score of 0.8277, and NN, which achieves the highest F1 score of 0.4923. Figure 2A shows that biopsy GG, percent positive core, PSA and age have a significant influence on LR predictions, while Fig. 2B shows that biopsy GG, PI-RADS, age, and percent positive core are impactful factors for NN. Biopsy GG, age, and percent positive core are identified as the common features that affect predictions in both models.

Table 3

Comparison of model performance through over-, under-, and no-sampling strategies where the highest value for each study is in bold font.
Strategies	Model	Training Accuracy	Testing Accuracy	Precision	Recall	F1
Over-sampling
	XGBoost	1	0.7698	0.4177	0.3819	0.3937
	NN	0.8729	0.6719	0.3076	0.4444	0.3598
	Naïve Bayes	0.6561	0.5660	0.2972	0.8014	0.4327
	RF	1	0.8037	0.5260	0.3986	0.4438
	SVM	0.6252	0.6047	0.2972	0.5235	0.3443
	LR	0.7002	0.7171	0.3891	0.6028	0.4601
Under-sampling
	XGBoost	0.8364	0.7169	0.3734	0.5222	0.4320
	NN	0.7466	0.6358	0.3152	0.6528	0.4198
	Naïve Bayes	0.6705	0.6234	0.3250	0.7583	0.4509
	RF	0.9529	0.7027	0.3727	0.6167	0.4592
	SVM	0.7768	0.6714	0.3526	0.7112	0.4680
	LR	0.6962	0.6262	0.3208	0.6946	0.4339
No-sampling
	XGBoost	0.8532	0.8226	0.4267	0.1764	0.2422
	NN	0.8279	0.8107	0.4083	0.6306	0.4923
	Naïve Bayes	0.7194	0.7317	0.6353	0.3816	0.3975
	RF	0.9060	0.8229	0.6167	0.2392	0.3161
	SVM	0.7106	0.6466	0.3889	0.5639	0.4041
	LR	0.8335	0.8277	0.6533	0.3527	0.4278

The decision-making process between active surveillance and surgical intervention, regarding GG upgrading in PCa, especially towards clinically significant PCa, such as ISUP GG 2 to 3, is of significant importance(6, 7). Despite extensive research into potential factors, the exact features that contribute to upgrading remain controversial. Therefore, our study aims to integrate both potential characteristics and time factors to predict upgrade, considering a maximum waiting time for RP of approximately 3 to 4 months, as depicted in Table 1. Notably, the median surgical waiting time in Class 0 (Non-upgrading) is 2 weeks longer than Class 1 (Upgrading).

Contrary to expectations, our findings revealed no specific features that demonstrated a strong association with GG upgrade. Both the Pearson correlation and the PCA failed to identify the distinct factors that contributed significantly to the upgrading. The surgical waiting time and most features do not show correlations, suggesting that the threshold of 3–4 months did not significantly impact upgrading. This is consistent with previous studies(13, 14) and existing guidelines(29), which support the safe delay of RP at around 3 months. However, three features exhibit weak correlations, including negative correlations with biopsy GG and PI-RADS, and a positive correlation with a percent positive core. These weak correlations may be attributed to the rarity of the upgrading events compared to the more prevalent non-upgrading events, influencing the degree of associations in the data. This is consistent with previous studies on percent positive core (9, 10, 12, 30) and PI-RADS(12). As also emphasized in machine learning application literature(30, 31), there's a consistent recognition of imaging characteristics in predicting GG upgrades. Furthermore, despite the comprehensive nature of PCA, it did not offer additional insights beyond those provided by Pearson's correlation. The scree plot in Fig. 1A, where the 7 PCs required to obtain more than 80% explained variance, indicates that individual features alone are insufficient for accurate predictions. The limited discriminatory power illustrated in Fig. 1B poses the challenge of identifying distinctive features, making model training in our dataset a challenging task.

Therefore, we then employed 10-fold and 3-fold cross-validation within each fold, assisting the model in learning challenging cases. Addressing data imbalance during the preprocessing stage, we implemented both over-sampling and under-sampling (32). It is essential to note that the statistical significance for comparing model performance among resampling circumstances is not applicable, as the evaluation relies primarily on performance metrics such as accuracy, precision, recall, and F1 score. Interestingly, both over-sampling and under-sampling strategies contributed to the enhancement of model performance for most models, except for NN and LR. Regarding NN, the decrease in performance is attributed to the overfitting of models resulting from resampling techniques, as indicated by high testing accuracy but low training accuracy. For LR, as discussed by van den Goorbergh et al. (33), data imbalance can have an adverse effect on the parameter tuning process. Despite a slight decline in model performance, resulting from a slight imbalance in our study, LR remains the best model in accuracy in all three strategies, while NN performs the best in terms of the F1 score. Furthermore, among ML models, those achieving testing accuracy exceeding our baseline of 82.1% included XGBoost, RF, and LR when trained on data without applying any sampling methods.

Considered the best model in the present study, we perform an explanatory analysis on LR and NN using the Explained add-on library in Orange Data Mining to assess the features influencing the model output, shown in Fig. 2. Consistent with previous studies(8–10, 12), the main common contributors to the predictions of LR and NN were biopsy GG, age and percent positive core; however, these characteristics did not show a high impact on the model, supported by weak correlation scores in Fig. 1 and Table 2. Regarding the clinical application of our prediction model, we want the model to accurately identify all cases labeled 1 (upgrade). Despite the challenges, our empirical goal is to attain a minimum precision of 0.8 for label 1, corresponding to an F1 score threshold of 0.89. However, our model performance does not meet this F1 threshold, indicating that this model does not meet the required performance for clinical use. This highlights the need to explore additional features to enhance predictive capabilities.

Concerning the time waiting for surgery, both our study and existing research prove that analyzing at a specific time point does not provide clear predictions. We propose exploring alternative models, such as the Cox proportional hazards model, to enhance time-associated predictions, considering not only one time point but the entire duration. Given the weak to no correlations of our features with the target feature and the failure to reach the F1 threshold in model performance, there arises a need for further exploration of supplementary features that could serve as indicators for this problem. In clinical practice, even with a weak correlation, clinicians must be attentive if a patient with a low GG also has high PSA levels, is older, or has a high percentage of positive core, as these suggest an increased risk of upgrading. It is worth suggesting for future studies that predicting upgrades may not rely solely on specific individual features. The physician's experience of the physician and the characteristics of the patient are likely important factors that influence the decision about the next step in individualized treatment.

This study has limitations. First, being a single-center study conducted in Thailand, its generalizability may be limited, particularly in terms of racial or genetic factors. Second, with surgical waiting time of only around 3–4 months, its long-term relevance to other settings may be limited. Third, as a retrospective chart review, the data and time-dependent factors could influence decisions in practice, potentially affecting the analysis.

In conclusion, there are no features strongly associated with GG upgrading, which might be due to the rarity of the event and the disproportion of our dataset. GG at biopsy, percent positive core, and PI-RADS appear to have a weak correlation with GG upgrading, similar to previous research. Regarding time-related features, the duration awaiting surgery approximately 3–4 months does not significantly affect the prediction of upgrade. Notably, incorporating a specific time-point factor into the analysis failed to improve the model prediction. We recommend further research exploring alternative time-associated model analyzes and additional features that could help improve prediction.

Conflict of interest

All authors declare no personal or professional conflicts of interest, and no financial support from the companies that produce and/or distribute the drug, devices, or materials described in this report.

Author Contribution

K.M. Writing – original draft, Formal analysis, Data curation. P.A. Writing – Conceptualization, Resources, Data curation. K.L. Resources, Data curation. A.H. Writing – review & editing, Methodology, Conceptualization, Formal analysis.S.S. Resources, writing – review & editing, Conceptualization, Project administration.

Acknowledgement

The author would like to express gratitude to Miss Jitsiri Chaiyatho and the coordinators at Siriraj Hospital for their valuable contributions to this study.

Data Availability

The data from this study are not available in a publicly accessible repository. However, they can be made available upon request for academic purposes. For further details, please contact S.S., the corresponding author.

Observatory, T. G. C. Thailand fact sheet [Internet]. Lyon, France: International Agency for Research on Cancer, World Health Organization; 2020 [cited 2023 Sep 9]. https://gco.iarc.fr/today/data/factsheets/populations/764-Prostate-fact-sheets.pdf
Observatory, T. G. C. & World Health Organization. Prostate cancer fact sheet [Internet]. Lyon, France: International Agency for Research on Cancer, ; [cited 2023 Sep 9]. (2020). https://gco.iarc.fr/today/data/factsheets/cancers/27-Prostate-fact-sheet.pdf
Thailand NCIo. Hospital-based cancer registry 2021 [Internet]. Bangkok, Thailand: Ministry of Public Health; 2022 [cited 2023 Sep 9]. https://www.nci.go.th/e_book/hosbased_2564/index.html
Mottet, N. et al. EAU-EANM-ESTRO-ESUR-SIOG Guidelines on Prostate Cancer-2020 Update. Part 1: Screening, Diagnosis, and Local Treatment with Curative Intent. Eur. Urol. 79 (2), 243–262 (2021).
Ahmed, H. U. et al. Diagnostic accuracy of multi-parametric MRI and TRUS biopsy in prostate cancer (PROMIS): a paired validating confirmatory study. Lancet. 389 (10071), 815–822 (2017).
Bakavicius, A. et al. The Impact of Prostate Cancer Upgrading and Upstaging on Biochemical Recurrence and Cancer-Specific Survival. Med. (Kaunas) ;56(2). (2020).
Kovac, E., Vertosick, E. A., Sjoberg, D. D., Vickers, A. J. & Stephenson, A. J. Effects of pathological upstaging or upgrading on metastasis and cancer-specific mortality in men with clinical low-risk prostate cancer. BJU Int. 122 (6), 1003–1009 (2018).
Caster, J. M., Falchook, A. D., Hendrix, L. H. & Chen, R. C. Risk of Pathologic Upgrading or Locally Advanced Disease in Early Prostate Cancer Patients Based on Biopsy Gleason Score and PSA: A Population-Based Study of Modern Patients. Int. J. Radiat. Oncol. Biol. Phys. 92 (2), 244–251 (2015).
Dinh, K. T. et al. Incidence and Predictors of Upgrading and Up Staging among 10,000 Contemporary Patients with Low Risk Prostate Cancer. J. Urol. 194 (2), 343–349 (2015).
Epstein, J. I., Feng, Z., Trock, B. J. & Pierorazio, P. M. Upgrading and Downgrading of Prostate Cancer from Biopsy to Radical Prostatectomy: Incidence and Predictive Factors Using the Modified Gleason Grading System and Factoring in Tertiary Grades. Eur. Urol. 61 (5), 1019–1024 (2012).
Brassetti, A. et al. Prostate-specific Antigen Density Is a Good Predictor of Upstaging and Upgrading, According to the New Grading System: The Keys We Are Seeking May Be Already in Our Pocket. Urology. 111, 129–135 (2018).
Song, W. et al. Role of PI-RADS Version 2 for Prediction of Upgrading in Biopsy-Proven Prostate Cancer With Gleason Score 6. Clin. Genitourin. Cancer. 16 (4), 281–287 (2018).
Filippou, P. et al. Immediate versus delayed radical prostatectomy: updated outcomes following active surveillance of prostate cancer. Eur. Urol. 68 (3), 458–463 (2015).
Weiner, A. B., Patel, S. G. & Eggener, S. E. Pathologic outcomes for low-risk prostate cancer after delayed radical prostatectomy in the United States. Urol. Oncol. 33 (4), 164e11–164e17 (2015).
Beam, A. L. & Kohane, I. S. Big Data and Machine Learning in Health Care. JAMA. 319 (13), 1317–1318 (2018).
Jordan, M. I. & Mitchell, T. M. Machine learning: Trends, perspectives, and prospects. Science. 349 (6245), 255–260 (2015).
Sidey-Gibbons, J. A. M. & Sidey-Gibbons, C. J. Machine learning in medicine: a practical introduction. BMC Med. Res. Methodol. 19 (1), 64 (2019).
Riley, R. D. et al. Calculating the sample size required for developing a clinical prediction model. BMJ. 368, m441 (2020).
Demšar, J. et al. Orange: Data Mining Toolbox in Python. J. Mach. Learn. Res. 14, 2349–2353 (2013).
Schober, P., Boer, C. & Schwarte, L. Correlation Coefficients: Appropriate Use and Interpretation. Anesth. Analgesia. 126, 1 (2018).
Jolliffe, I. T. Choosing a Subset of Principal Components or Variables. In: (ed Jolliffe, I. T.) Principal Component Analysis. New York, NY: Springer New York; 92–114. (1986).
XGBoost d. XGBoost Documentation [Internet]. xgboost developers; 2023 [cited 2024 Jan 9]. https://xgboost.readthedocs.io/en/stable/index.html
Cortes, C. & Vapnik, V. Support-vector networks. Mach. Learn. 20 (3), 273–297 (1995).
Mitchell, T. Machine Learning. New York1432 (McGraw-Hill Professional, 1997 Mar).
Developers Gf. Imbalanced Data | Machine Learning2022 2024 Jan 9. https://developers.google.com/machine-learning/data-prep/construct/sampling-splitting/imbalanced-data
Kulkarni, A., Chong, D. & Batarseh, F. Foundations of data imbalance and solutions for a data democracy2021.
The pandas development. t. pandas-dev/pandas: Pandas. Zenodo; (2023).
Pedregosa, F. et al. Scikit-learn: Machine learning in Python. J. Mach. Learn. Res. 12, 2825–2830 (2011).
Guidelines, E. & EAU - EANM - ESTRO - ESUR - ISUP - SIOG Guidelines on Prostate Cancer [Internet]. EAU Guidelines Office: Arnhem, The Netherlands; 2023 [cited 2024 Jan 25]. http://uroweb.org/guidelines/compilations-of-all-guidelines/
Liu, H. et al. Predicting Prostate Cancer Upgrading of Biopsy Gleason Grade Group at Radical Prostatectomy Using Machine Learning-Assisted Decision-Support Models. Cancer Manage. Res. 12 (null), 13099–13110 (2020).
Ozbozduman, K. et al. Machine learning prediction of Gleason grade group upgrade between in-bore biopsy and radical prostatectomy pathology. Sci. Rep. 14 (1), 5849 (2024).
Branco, P., Torgo, L. & Ribeiro, R. A Survey of Predictive Modelling under Imbalanced Distributions2015.
van den Goorbergh, R., van Smeden, M., Timmerman, D. & Van Calster, B. The harm of class imbalance corrections for risk prediction models: illustration and simulation using logistic regression. J. Am. Med. Inform. Assoc. 29 (9), 1525–1534 (2022).

No competing interests reported.

Download PDF

Reviewers agreed at journal
20 Nov, 2024
Reviewers invited by journal
03 Sep, 2024
Editor assigned by journal
03 Sep, 2024
Editor invited by journal
02 Sep, 2024
Submission checks completed at journal
02 Sep, 2024
First submitted to journal
24 Aug, 2024

You are reading this latest preprint version

Machine Learning-Based Prediction of Gleason Grade Group upgrading in Patients with Localized Prostate Cancer Awaiting Surgery

Status:

Version 1

Abstract

Figures

Introduction

Materials and Methods

Patient selection, Data preparation, and Study parameters

Machine Learning Analysis

Statistical Analysis

Results

Patient Characteristics

Data Exploration

Model Performances Across Sampling Strategies

Discussion

Conclusions

Declarations

Conflict of interest

Author Contribution

Acknowledgement

Data Availability

References

Additional Declarations

Status:

Version 1