Risk prediction models for biochemical recurrence of Chinese prostate cancer patients after radical prostatectomy based on magnetic resonance imaging examination: a systematic review

Abstract

Background

Biochemical recurrence (BCR) after radical prostatectomy (RP) affects the prognosis of patients and early accurate prediction is crucial. Magnetic resonance imaging (MRI) is of great value in the assessment of prostate cancer (PCa). However, there is a lack of a systematic summary of the current research status for the construction of a postoperative BCR prediction model applicable to Chinese PCa patients based on MRI features. This study aimed to systematically evaluate the predictive performance and clinical applicability of the available models.

Methods

A standardized search of relevant literature in the PubMed, Cochrane Library, Embase, Web of Science, CINAHL, CNKI, VIP, Wanfang Data, and CBM databases was performed, with the search time restricted to the establishment of the database to 11 September 2024. Studies that developed and/or validated prediction models based on MRI examination to identify and/or predict BCR in patients after RP in China were included. Two researchers independently screened the literature and used the prediction model risk of bias assessment tool to assess the quality of research on the prediction models and performed descriptive analyses of predictor variables for modeling.

Results

A total of 17 studies were included, and 41 prediction models for BCR risk in Chinese patients after RP based on MRI examination were constructed, with the area under the receiver operating characteristic curve (AUC) or concordance index (C-index) of the cases ranging from 0.610 to 0.982. A total of 36 prediction models had good predictive performance, eight studies performed model calibration, two studies performed internal validation, two studies performed external validation, and seven studies conducted both internal and external validation. The results of the quality assessment revealed that all 17 studies were at high risk of bias. The most frequent predictors were prostate-specific antigen (PSA) level, MRI image features, and Gleason score.

Conclusions

At present, a prediction model based on MRI examination for the risk of BCR in Chinese patients after RP is still in the development stage, and the overall quality of research needs to be further improved. In the future, the study design and reporting process should be improved, and the existing model should be validated to provide a basis for the development of effective prevention strategies.

Introduction

Prostate cancer (PCa) is the most common urological malignant tumor in elderly men and has the highest incidence in European and American countries (1); however, in recent years, with improvements in living standards, population aging, and lifestyle changes, the incidence of PCa in China has shown a significant upward trend (2). Radical prostatectomy (RP) for PCa is one of the main treatment modalities and can effectively remove tumor lesions and improve patient survival, but biochemical recurrence (BCR) occurs in approximately 27–53% of patients after surgery (3,4). BCR tends to be highly correlated with the occurrence of local recurrence, distant metastasis, PCa-specific and overall mortality rates in patients, among other factors (5,6). As BCR presents significant heterogeneity, it is difficult to predict its occurrence comprehensively and accurately via analysis of clinical and pathological information alone (7). Therefore, more accurate predictive tools are needed to identify high-risk patients so that appropriate treatment plans can be developed early.

As a core imaging tool for PCa evaluation, magnetic resonance imaging (MRI) has the unique advantage of multisequence and multiparameter imaging. Its excellent soft-tissue resolution does not only clearly display the anatomical subdivisions of the prostate and the relationship between the surrounding tissues but also visually reveals the extent of tumor invasion of the peritoneum, seminal vesicles, urethral sphincter, and other structures through T2-weighted imaging (T2WI), which can provide a key basis for preoperative clinical staging (8). Compared with ultrasound and computed tomography (CT), the functional imaging techniques of MRI can capture changes in the tumor microenvironment, such as diffusion-weighted imaging (DWI), which reflects the density of tumor cells by measuring the apparent diffusion coefficient (ADC), and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), which assesses the state of tumor angiogenesis through quantitative analysis of hemodynamic parameters (9). In addition, MRI multiparameter fusion technology can integrate morphological features and functional parameters to form a multidimensional portrayal of tumor biological behaviors, providing rich imaging features and variables for prediction models (10,11). On the basis of the morphological, functional, and molecular biological features of the tumor obtained from MRI, a prediction model for BCR of PCa after RP can be constructed, which is expected to provide a more accurate recurrence risk assessment tool for clinicians.

To date, a variety of BCR risk prediction models based on MRI examination of patients after RP have been developed both within China and internationally. However, there are large differences in study design, including population, MRI examination sequences, and parameter analysis, and relatively few studies have been conducted on the Chinese population. Existing studies have found that foreign prediction models are not fully applicable to the Chinese population because of the large racial differences between Eastern and Western populations in terms of tumor biological characteristics, and the Chinese population has higher prostate-specific antigen (PSA) levels and Gleason scores (12,13). Therefore, in this study, we searched for studies on BCR risk prediction models for post-RP patients in China on the basis of MRI examinations; used a systematic evaluation method to summarize the current research status of BCR prediction models for post-RP patients in China on the basis of MRI examinations; and analyzed the construction method, inclusion indicators, and prediction efficacy of each prediction model. We aimed to provide scientific evidence for the future development, application, optimization, and personalized prevention and treatment of this type of risk prediction model in China. We present this article in accordance with the PRISMA reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2843/rc).

Methods

Literature search

A standardized search of relevant literature in the PubMed, Cochrane Library, Embase, Web of Science, CINAHL, CNKI, VIP, Wanfang Data, and CBM databases was performed, with the search time restricted to the establishment of the database to 11 September 2024. Searches were conducted via a combination of subject terms, free words, and literature tracing methods. The search terms included “radical prostatectomy/prostatic carcinoma/Prostatic Neoplasm*/prostatectomy/Prostate Neoplasm*/Prostate Cancer*/Cancer of Prostate/Prostatic Cancer*/Prostatectom*/suprapubic prostatectom*/retropubic prostatectom*/Prostatic tumour” “Recurrence/Biochemical Recurrence/Recurrence*/Relapse*/Recrudescence*/biochemical failure/BCR” “Magnetic Resonance Imaging/Multiparametric Magnetic Resonance Imaging/Radiomics/Diagnostic Imaging/mpMRI/magnetic resonance/MRI/Magnetic Resonance Images/imageological diagnosis/Imaging studies/Medical Imaging” and “Risk prediction Model/Nomogram/Model Building/Risk Assessment/prognos*/predict*/risk factor*”. The detailed search strategy can be found in the Appendix 1. All data retrieval and collection were completed on the same day, 12 September 2024, to avoid bias due to database updates. This study was reviewed via the PROSPERO platform (CRD42024586153).

Inclusion and exclusion criteria

Inclusion criteria: (I) the cases were Chinese RP patients; (II) the study was a construction and/or validation study of a BCR risk prediction model (excluding prognostic and progression models) for Chinese post-RP patients based on MRI examinations; (III) prospective, retrospective, and cross-sectional studies; (IV) Chinese and English literature.

Exclusion criteria: (I) only risk factors were analyzed without constructing prediction models; (II) model predictors <2; (III) studies based on systematic reviews or meta-analyses; (IV) inability to access the full text, incomplete information, and indicators that could not be extracted or repeatedly published.

Literature screening and data extraction

Literature management was performed via NoteExpress software (Aegean Software, Beijing, China) to screen literature on the basis of title, abstract, and full text. Forms were developed for data extraction on the basis of critical appraisal and data extraction for systematic reviews of prediction modeling studies (CHARMS) (14). Literature screening and data extraction were independently performed and cross-checked by two investigators, and disagreements were resolved via discussion with a third investigator.

Model quality evaluation methods

The prediction model risk of bias assessment tool (PROBAST) (15) was used by two researchers to independently evaluate the risk of bias and applicability of the included studies. If there were any doubts, a third researcher arbitrated. The extracted information of each study was entered into Microsoft Excel (Microsoft Corp., Redmond, WA, USA). On account of the heterogeneity and paucity of studies in this review, a qualitative synthesis of the results used a narrative approach.

Results

Literature screening process and results

A total of 4,529 relevant studies were initially retrieved in this study. After the software automatically removed duplicates and manually removed literature that did not meet the inclusion exclusion criteria, 17 studies were finally included (16-32), including seven Chinese studies (16,21-23,26,27,30) and 10 English-language studies (17-20,24,25,28,29,31,32). The retrieval process is shown in Figure 1.

Figure 1.

PRISMA 2020 flow diagram of the selection process of the included studies. PRISMA, Preferred Reporting Items for Systematic Reviews and Meta-Analyses.

Basic characteristics of the included studies

The studies included in this paper were published within the last 10 years, and the basic characteristics are shown in Table 1.

Table 1. Basic characteristics of the included studies.

Included studies	Publication year	Country	Study type	Research participants	Source of data	Definition of outcome (predictive results diagnostic methods)
Yan XX (16)	2016	China	Retrospective	RP patients	The Second Hospital of Tianjin Medical University	PSA examination after two consecutive surgeries reached above 0.2 ng/mL
Zhang YD (17)	2016	China	Retrospective	RP patients	The First Affiliated Hospital of Nanjing Medical University	Postoperative detectable PSA levels (≥0.2 ng/mL)
Zhang YD (18)	2017	China	Retrospective	RP patients	The First Affiliated Hospital of Nanjing Medical University	Postoperative detectable PSA levels (≥0.2 ng/mL)
Xia HZ (19)	2021	China	Retrospective	RP patients	Peking University Third Hospital	Twice measured postoperative PSA values >0.2 ng/mL
Yan Y (20)	2021	China	Retrospective	RP patients	Peking University Third Hospital, Beijing Friendship Hospital, Peking University People’s Hospital	RP with consecutive total PSA readings over 0.2 ng/mL
Chen ZH (21)	2022	China	Retrospective	RP patients	Huadong Hospital Affiliated to Fudan University	PSA >0.2 ng/mL on 2 consecutive occasions, PSA >0.4 ng/mL, or receiving postoperative salvage therapy
Wang KX (22)	2022	China	Retrospective	RP patients under laparoscopy	Qingdao University Affiliated Hospital	PSA levels increased during the follow-up after RP surgery, two consecutive serum PSA ≥0.2 ng/mL, and routine imaging examinations did not find recurrence or metastasis
Yang H (23)	2022	China	Retrospective	RP patients	Taizhou Hospital of Chinese Medicine	PSA monitoring for consecutive 2 times ≥0.2 ng/mL
Chen Z (24)	2022	China	Retrospective	RP patients	Huadong Hospital Affiliated to Fudan University	PSA >0.2 ng/mL twice in a row, PSA >0.4 ng/mL or postoperative salvage therapy
Wang Y (25)	2022	China	Retrospective	RP patients	Xiangyang First People’s Hospital and Jiangsu Provincial Hospital of Chinese Medicine	PSA level increased ≥0.2 ng/mL in two or more consecutive follow-up visits
Ji YY (26)	2023	China	Retrospective	RP patients	The First Affiliated Hospital of Soochow University	PSA after surgery has increased twice in a row and the second PSA ≥0.2 ng/mL
Zhang YF (27)	2023	China	Retrospective	RP patients	The First Affiliated Hospital of Soochow University	Continuous PSA levels ≥0.2 ng/mL for two consecutive times postoperatively
An P (28)	2023	China	Retrospective	RP patients	Xiangyang First People’s Hospital	PSA level increased twice or more (≥0.2 ng/mL)
Hou Y (29)	2023	China	Retrospective	RP patients	The First Affiliated Hospital of Nanjing Medical University	Three consecutive PSA increases after surgery >0.1 ng/mL, at least 6 weeks, and finally PSA >0.2 ng/mL or at least 6 weeks after surgery PSA ≥0.4 ng/mL, or secondary treatment due to the increase in PSA levels, reference to previous reports and correlated with the likelihood of subsequent PSA progression
Liu W (30)	2024	China	Retrospective	RP patients	Beijing Hospital	Postoperative RP, continuous follow-up 2 times, PSA values rose to above 0.2 ng/mL, and there is a rise
Hu C (31)	2024	China	Retrospective	RP patients	The First Affiliated Hospital of Soochow University	The total PSA level measured twice consecutively after RP is 0.2 ng/mL or higher
Xia J (32)	2024	China	Retrospective	RP patients	Shaoxing University Affiliated Hospital	PSA continuous 2 times >0.2 ng/mL

PSA, prostate-specific antigen; RP, radical prostatectomy.

Construction of the risk prediction model

The 17 included studies constructed 41 BCR risk prediction models for Chinese patients after RP on the basis of MRI examination. With respect to model construction, one study (18) used multiple least squares regression to build prediction models, and seven studies (17,22,23,28-31) used machine learning algorithms to build models. Zhang (27) selected the optimal model by comparing algorithms such as logistic regression, random forest, support vector machine, least absolute shrinkage and selection operator (LASSO) regression and established the model via logistic regression or Cox proportional hazards regression (Cox regression) for other terms. In terms of variable treatment, six studies (19,24,26,29-31) maintained the continuity of variables, seven studies (16-18,20-22,32) converted continuous variables into categorical variables, and the remaining studies (23,25,27,28) did not mention continuous variables (see Table 2).

Table 2. Model establishment.

Included studies	Number of models	Modeling method	Variable selection	Sample size		Method for handling missing values	Method for handling continuous variables	Number of predictors	Included predictors
				Negative event	Positive event
Yan XX (16)	2	Cox	Single + multiple	28	87	Exclude	Categorical variable	3	Model 1: MRI examination + Gleason score + clinical stage
								2	Model 2: Gleason score + clinical stage
Zhang YD (17)	3	SVM	Single + multiple	144	61	Exclude	Categorical variable	8	MRI model: age, PI-RADS score, tumor location, MTD, MR visibility, tumor ADCS, DCE type, tumor T staging detected by MR
								4	SVM’s D’Amico model: age, PSA level, Gleason score, TNM stage
								11	SVM’s D’Amico + MR model: age, PI-RADS score, tumor location, MTD, MR visibility, tumor ADCs, DCE type, tumor T staging detected by MR, PSA level, Gleason score, TNM stage
Zhang YD (18)	3	PLS, Cox	Single + multiple	153	52	Exclude	Categorical variable	8	MRI model: age, PSA level, tumor location, MTD, tumor ADCs, PI-RADS score, DCE type, tumor T staging detected by MRI
								9	Improved D’Amico: TNM staging, Gleason score, PSA level, tumor location, MTD, tumor ADCs, PI-RADS score, DCE type, tumor T staging detected by MRI
								11	Improved CAPRA score: age, PSA level, Gleason score, percentage of positive cores in biopsy for cancer, TNM staging, tumor location, MTD, tumor ADCs, PI-RADS score, DCE type, tumor T staging detected by MRI
Xia HZ (19)	2	Cox	Single + multiple	237	100	Exclude	Continuous variable	5	New nomogram: PSA level, MTD, Gleason score, SM, SVI
								3	Basic model Gleason score, SM, SVI
Yan Y (20)	2	Cox	Single	372	113	Exclude	Categorical variable		DRS-BCR: for deep imaging genomic features, specific predictive factors are not mentioned
									CS is a clinical feature, and the specific prediction factor is not mentioned
Chen ZH (21)	1	Cox	Single + multiple	147	63	Exclude	Categorical variable	3	MTD, Gleason score, SVI
Wang KX (22)	1	BN	Single	166	43	Exclude	Categorical variable	10	PI-RADS score, Gleason score, capsule invasion, SVI, lymphovascular and perineural invasion, SM, clinical stage, pathological stage, surgical method, treatment method
Yang H (23)	3	LASSO	Single + multiple	104	65	Exclude	Not mentioned	6	Combined model: MTD, PSA level, treatment method, elastography grading, grayscale region size matrix features - small region high grayscale emphasis range, grayscale travel matrix features-travel variance
								3	General information model: lymph node metastasis, PSA level, treatment method
								3	Magnetic resonance oncology model: grayscale region size matrix features-small region high grayscale emphasis range, grayscale travel matrix features-travel variance, neighborhood gray difference matrix feature-signal-to-noise ratio contrast
Chen Z (24)	1	Cox, LR	Single + multiple	215	89	Exclude	Continuous variable	3	MTD, SVI, Gleason score
Wang Y (25)	4	LR	Single + multiple	100	59	Exclude	Not mentioned	7	General data model: MTD, PSA level, treatment method, lymph node metastasis, T staging, Gleason score, distant metastasis
								4	CEUS model: PI, TTP, AT, elastic imaging level
								3	MRI radiomics model: small region high gray emphasis range, track variance, signal-to-noise ratio contrast
								14	Combined model: MTD, PSA level, treatment method, lymph node metastasis, T staging, Gleason score, distant metastasis, PI, TTP, AT, elastography grade, small region high gray emphasis range, track variance, signal-to-noise ratio contrast
Ji YY (26)	2	Cox	Single	B: 99	B: 50	Exclude	Continuous variable	26	Imaging group learning model: one first-order statistical quantitative characteristics, two shape characteristics, two texture characteristics, ten texture characteristics based on wavelet conversion, three first-order statistical characteristics based on wavelet conversion, two filter-based conversion of filter-based conversion first-order statistical characteristics, the Radscore composed of 6 filter-based texture characteristics
				V: 42	V: 21			4	Clinical-imaging biomarker model construction: Radscore, PSA level, Gleason score, clinical T staging
Zhang YF (27)	3	SVM, LR, LASSO, RF	ANOVA, RFE	B: 45	B: 35	Exclude	Not mentioned	13	Model-T2 established by T2WI sequence: 13 mpMRI features
				V: 19	V: 15			9	Model A + D built with DWI and ADC sequences: 9 mpMRI features
								12	Model-M model built jointly: 12 mpMRI features
An P (28)	4	LASSO	Single + multiple	B: 91	B: 54	Exclude	Not mentioned	6	General clinical model: MTD, TNM staging, lymph node transfer, distant transfer, Gleason score, PSA level
								3	CEUS model: PI, AT, elastography grade
				V: 38	V: 23			2	MRI radiation model: Radscore 1, Radscore 2
								11	Integrated model: MTD, TNM staging, lymph node transfer, distant metastasis, Gleason score, PSA level, PI, AT, elastography grade, Radscore 1, Radscore 2
Hou Y (29)	2	LASSO, Cox	Not mentioned	B: 326	B: 137	Not mentioned	Continuous variable	9	Radiomics model: Radscore composed of 1 feature from T2WI, 5 features from DWI, and 3 features from ADC
				V: 82	V: 34			3	M5 model: Radscore, ECE, PSA
Liu W (30)	3	LASSO, Cox	Single + multiple	B: 98	B: 24	Exclude	Continuous variable	3	Clinical model PSA level, pathological staging, SM
				V: 42	V: 11			5	Imaging model: 5 mpMRI features
								8	Imaging-clinical combined model: PSA level, pathological stage, SM, five mpMRI features
Hu C (31)	4	LASSO, Cox	Not mentioned	B: 195	B: 59	Exclude	Continuous variable	3	Clinical features: PSA level, T staging, Gleason score
								12	Imaging pathology: 12 imaging pathology features
				V: 84	V: 25			4	Pathology groupomics: 4 pathology groupomics features
								19	Combination model: PSA level, T staging, Gleason score, 12 imaging groupomics features, four pathology groupomics features
Xia J (32)	1	LR	Single + multiple	50	52	Exclude	Categorical variable	2	PSA level, PI-RADS score

ADC, apparent diffusion coefficient; ANOVA, analysis of variance; AT, arrival time; B, building set; BN, Bayesian network; CEUS, contrast-enhanced ultrasound; Cox, Cox proportional hazards regression; D’Amico, D’Amico scheme; DCE, dynamic contrast enhancement; DRS-BCR, deep-radiomic signature-biochemical recurrence; DWI, diffusion-weighted imaging; ECE, extracapsular extension; LASSO, least absolute shrinkage and selection operator; LR, logistic regression; mpMRI, multiparametric magnetic resonance imaging; MR, magnetic resonance; MRI, magnetic resonance imaging; MTD, maximum tumor diameter; PI, peak intensity; PI-RADS, prostate imaging-reporting and data system; PLS, partial least square regression analysis; PSA, prostate-specific antigen; RF, random forest; RFE, recursive feature elimination; Single + multiple, single factor analysis plus multiple factor analysis; SM, surgical margin; SVM, support vector machine; SVI, seminal vesicle invasion; T2WI, T2-weighted imaging; TNM, tumor-node-metastasis; TTP, peak time; V, validation set.

Model performance and predictors

The discriminative power of the models involved in this study was evaluated mainly by area under the receiver operating characteristic curve (AUC) or concordance index (C-index). Two studies (20,30) used both the AUC and the C-index to evaluate model discrimination, three studies (19,29,31) used the C-index, and the remaining studies used the AUC to assess discrimination. Except for four studies (19,20,28,31), the AUC/C-index (range, 0.610 to 0.982) of 36 models constructed in the other studies was greater than 0.7, indicating good discrimination. Eight studies (19,20,23,25,26,28,29,31) involved model calibration, two studies (18,19) involved internal validation, two studies (20,23) involved external validation, and seven studies (25-31) involved combined internal and external validation for model evaluation. Among the studies, 2–26 predictors were included in the model. The most frequent predictors were PSA level (n=20), MRI image features (n=18), Gleason score (n=17), and so on. Six (19,25,26,28,29,31) prediction models were finally presented in the form of nomograms, seven studies (16,18,20,23,27,30,32) were not reported, one study (22) was a Bayesian network diagram, and the remaining three studies (17,21,24) were risk scoring formulas or scoring criteria. Model performance and presentation form are shown in Table 3.

Table 3. Model performance and presentation form.

Included studies	Model performance		Model validation	Model presentation
	Discrimination	Calibration
Yan XX (16)	Model 1: AUC =0.806	Not mentioned	Not mentioned	Not mentioned
	Model 2: AUC =0.763
Zhang YD (17)	LR MRI model: AUC =0.886 (0.834–0.926)	Not mentioned	Not mentioned	Risk scoring formula
	SVM MRI model: AUC =0.959 (0.922–0.982)
	SVM D’Amico model: AUC =0.859 (0.804–0.903)
	SVM D’Amico + MRI: AUC =0.970 (0.936–0.988)
Zhang YD (18)	MRI model: AUC =0.909 (0.861–0.944)	Not mentioned	Internal validation	Not mentioned
	D’Amico: AUC =0.793 (0.731–0.846)
	CAPRA score: AUC =0.809 (0.748–0.860)
	Improved D’Amico: AUC =0.901 (0.852–0.938)
	Improved CAPRA score: AUC =0.894 (0.843–0.932)
Xia HZ (19)	New Normotree: C =0.760 (0.710–0.810)	Calibration curve, DCA	Internal validation	Nomogram
	CAPRA-S model: C =0.700 (0.640–0.750)
	Basic model: C =0.660 (0.640–0.710)
Yan Y (20)	DRS-BCR: 3-year AUC =0.840, 5-year AUC =0.830	H-L test, calibration curve, DCA	External validation	Not mentioned
	CS: C =0.693 (0.634–0.752)
Chen ZH (21)	AUC =0.692 (0.614–0.770)	Not mentioned	Not mentioned	Risk scoring criteria
Wang KX (22)	AUC =0.814	Not mentioned	Not mentioned	Bayesian network map
Yang H (23)	Combination model: AUC =0.070 (0.844–0.969)	DCA	External validation	Not mentioned
	General data model: AUC =0.792 (0.698–0.885)
	Ultrasound contrast enhancement model: AUC =0.741 (0.641–0.841)
	MRI omics model: AUC =0.729 (0.626–0.831)
Chen Z (24)	AUC =0.729	Not mentioned	Not mentioned	Risk scoring criteria
Wang Y (25)	Universal data model: AUC =0.769 (0.675–0.863)	Calibration curve, DCA	Internal + external	Nomogram
	CEUS model: AUC =0.783 (0.692–0.873)
	MRI radiomics model: AUC =0.730 (0.629–0.830)
	Combined model: AUC =0.914 (0.854–0.974)
Ji YY (26)	Radiomics model: AUC =0.824 (0.701–0.948), C =0.784 (0.660–0.891)	Calibration curve, DCA	Internal + external	Nomogram
	Clinical-radiomics model: AUC =0.841 (0.714–0.968), C =0.802 (0.637–0.912)
Zhang YF (27)	Model-T2: AUC =0.982	Not mentioned	Internal + external	Not mentioned
	Model A + D: AUC =0.940
	Model-M model: AUC =0.916
An P (28)	General clinical model AUC =0.740 (0.670–0.810)	Calibration curve, DCA	Internal + external	Nomogram
	CEUS model AUC =0.610 (0.530–0.690)
	MRI radiation model AUC =0.850 (0.780–0.910)
	Comprehensive model AUC =0.910 (0.850–0.950)
Hou Y (29)	Radiomics model: C =0.718 (0.687–0.749)	Calibration curve, DCA	Internal + external	Nomogram
	Cox-PH M5 model: C =0.812 (0.750–0.874)
	Cox-GBM M5 model: C =0.816 (0.756–0.876)
	Cox-DL M5 model: C =0.826 (0.776–0.882)
Liu W (30)	Clinical model: C =0.751 (0.655–0.846)	Not mentioned	Internal + external	Not mentioned
	Imaging model: C =0.764 (0.655–0.872)
	Imaging-clinical combined model: C =0.782 (0.679–0.874)
Hu C (31)	Clinical characteristics: C =0.670 (0.592–0.741)	H-L test, calibration curve, DCA	Internal + external	Nomogram
	Radiomics feature: C =0.837 (0.794–0.880)
	Pathomics labeling: C =0.690 (0.617–0.763)
	Combined model: C =0.870 (0.838–0.906)
Xia J (32)	AUC =0.921 (0.900–0.942)	Not mentioned	Not mentioned	Not mentioned

Data in parentheses are shown as 95% confindence interval. AUC, area under the receiver operating characteristic curve; C, concordance index; CAPRA, the Cancer of the Prostate Risk Assessment; CAPRA-S, the Cancer of Prostate Risk Assessment Score; CEUS, contrast-enhanced ultrasound; Cox, Cox proportional hazards regression; Cox-DL, deep learning based Cox model; Cox-GBM, Cox gradient boosting machine; Cox-PH, Cox proportional hazards; CS, clinical signature; D’Amico, D’Amico scheme; DCA, decision curve analysis; DRS-BCR, deep-radiomic signature-biochemical recurrence; H-L test, Hosmer-Lemeshow test; Internal + external, internal validation + external validation; LR, logistic regression; MRI, magnetic resonance imaging; SVM, support vector machine.

Literature quality evaluation

All 17 included studies were at high risk of bias (Table 4). (I) Regarding the study participants, 17 studies (16-32) had a high risk of bias because of their retrospective study design, which may bias the sample size. (II) Regarding predictor variables, one study (20) used multicenter samples, and the data collected by each center may be different, so it was rated as “high risk”. (III) Results: three studies were rated as having a high risk of bias, two studies (17,18) had unreasonable classification methods for the results, and one study (26) may not have used standard outcome definitions. (IV) Analysis: all studies were rated as having a high risk of bias, and 10 studies (17,18,22,23,25-28,30,31) had insufficient outcome events with <20 events per variable (EPV); eight studies (16-22,32) discretized the continuous variables in part or all; three studies (20,22,26) screened predictive variables via univariate analysis; one study (29) did not report the treatment method for missing data; nine studies (16-18,21,22,24,27,30,32) did not report model calibration tests; and eight studies (16,17,20-24,32) did not use internal validation or only used random split validation to evaluate model fit.

Table 4. Quality evaluation of the included studies.

Included studies	Bias risk assessment					Applicability evaluation				Overall evaluation
	Research object	Predictors	Results	Analysis		Research object	Predictors	Results		Bias risk	Applicability
Yan XX, 2016 (16)	H	L	L	H		L	L	L		H	L
Zhang YD, 2016 (17)	H	L	H	H		L	L	N		H	N
Zhang YD, 2017 (18)	H	L	H	H		L	L	N		H	N
Xia HZ, 2021 (19)	H	L	L	H		L	L	L		H	L
Yan Y, 2021 (20)	H	H	L	H		L	L	L		H	L
Chen ZH, 2022 (21)	H	L	L	H		L	L	L		H	L
Wang KX, 2022 (22)	H	L	L	H		H	L	L		H	H
Yang H, 2022 (23)	H	L	L	H		L	L	L		H	L
Chen Z, 2022 (24)	H	L	L	H		L	L	L		H	L
Wang Y, 2022 (25)	H	L	L	H		L	L	L		H	L
Ji YY, 2023 (26)	H	L	H	H		L	L	N		H	N
Zhang YF, 2023 (27)	H	L	L	H		L	L	L		H	L
An P, 2023 (28)	H	L	L	H		L	L	L		H	L
Hou Y, 2023 (29)	H	L	L	H		L	L	L		H	L
Liu W, 2024 (30)	H	L	L	H		L	L	L		H	L
Hu C, 2024 (31)	H	L	L	H		L	L	L		H	L
Xia J, 2024 (32)	H	L	L	H		L	L	L		H	L

H, high risk; L, low risk; N, not mentioned.

In evaluating applicability, one study (22) was rated as having high applicability risk because its participants were limited to a specific population. Three studies (17,18,26) were judged to be unclear in terms of outcome because it was not specified where the definitions of the reported outcome measures came from. The remaining studies had good applicability.

Discussion

The prediction model had good prediction performance, but the overall risk of bias was high

The prediction models included in this study were all at high risk of bias, mostly focusing on the areas of research participants, outcomes, and analysis, for following main reasons: (I) the literature included in this study utilized retrospective data for modeling; although this type of data is easy to obtain and use, there may be problems such as missing data and incomplete inclusion of predictors. Cohort studies have good authenticity, representativeness, and independence, and choosing prospective data or registry data as modeling data when optimizing prediction models in the future is recommended to reduce the risk of data bias (33). (II) Some of the study outcomes did not adopt standardized definitions, which affected the objectivity and accuracy of the models and increased the risk of bias. Future studies should adopt a unified standardized definition to provide a uniform benchmark for the development, validation, application, and interpretation of predictive models and to ensure that comparisons between different models are reliable and reproducible. Notably, previous guidelines defined BCR as undetectable PSA after RP but ≥2 subsequent detections of elevated PSA with elevated thresholds reaching PSA >0.2 ng/mL (34,35). However, the latest guidelines (36) use PSA levels falling below the detection level after RP and two subsequent detections of PSA >1 ng/mL as the most recent criterion for BCR, and the direct clinical application of the existing model may lead to heterogeneity in the predictors and alterations in the incidence of BCR. (III) Insufficient predictors corresponding to outcome events, EPV ≥20 cases in model development studies can reduce model overfitting (33). In this paper, most of the studies were constructed on the basis of MRI examination of BCR risk prediction models for Chinese post-RP patients have more candidate predictors, so there is EPV <20, which leads to an increased risk of model bias and a decrease in credibility. More studies with large sample sizes should be conducted in the future, which may allow researchers to capture more variability and potential confounders to more accurately estimate predictor effects and improve the stability and reliability of the model, which will help to ensure its generalizability. (IV) The poor handling of variables in some studies and the conversion of continuous data into categorical variables for model construction may have resulted in excessive loss of model performance, and the difference in the distribution of variables between the applied and developed populations can lead to poor extrapolation of the model. When it is necessary to convert continuous variables into categorical variables for operation, the nonlinear fit of continuous variables should be tested, or the categorical variables should be defined using commonly recognized criteria, clinical significance, and so on (33). (V) The use of univariate screening of predictor variables may have problems such as incomplete inclusion of predictors and ignoring predictors with missing data (15). In the future, the screening of variables can be combined with clinical practice to adopt new methods, such as LASSO regression and stepwise regression (37), and parameter estimation and variable selection can be carried out at the same time to improve the accuracy of screening. In addition to the statistical perspective, predictors should be comprehensively included on the basis of their clinical significance, accessibility, and cost of measurement (33). (VI) Improper handling of missing data will bias the associations between predictors and outcomes and may affect accuracy, thus increasing model bias. Future studies should pay attention to the use of statistical data processing methods such as multiple interpolation and weighting (38,39) to increase the accuracy of prediction models. (VII) To adequately measure model performance, both discrimination and calibration should be assessed. Calibration reflects the degree of agreement between the predicted risk and the actual occurrence risk, usually using calibration plots, calibration slopes, and calibration intercepts (40), and using only the Hosmer-Lemeshow goodness-of-fit test for calibration or failing to report information on the discrimination and calibration of the predictive model can lead to a greater risk of bias. Future studies should evaluate and report these findings in a timely manner after modeling to facilitate comparisons of developed risk prediction models and their clinical translation. (VIII) A lack of performance assessment in studies and overfitting of the models will, to a degree, inhibit the applicability of the models. Some of the studies included in this review were not internally validated, which may have led to a biased assessment of model performance. In addition, the internal and external validation of prediction models, as a core step from modeling to clinical application, is related to the overall stability and applicability of the model and needs to be strictly controlled by researchers. Although some progress has been made in predictive modeling research, most predictive model construction studies do not have enough support for clinical application, and further improvement of model validation is still needed.

The predictors of the model have certain guiding significance for clinical practice

The predictors ultimately included in the models incorporated in this study were not the same, but certain commonalities still existed. The PSA level, MRI image features, and Gleason score were the most frequently occurring predictors, and future modeling could focus on considering the above factors. The results of several studies have demonstrated that the higher the preoperative serum PSA level, the greater the risk of postoperative BCR (41,42). PSA, as the most sensitive marker of PCa, can mediate the process of protein hydrolysis, affect the structure of the basement membrane, and promote the migration and metastasis of PCa (43). In addition, higher preoperative PSA levels in patients with PCa represent larger tumor tissues and more invasive epithelial cells and ducts, which in turn allows the PSA to enter the bloodstream, where its level is further elevated and the risk of postoperative BCR is increased (44). Meanwhile, the incorporation of a multifactorial prediction model of MRI image features can improve the prediction performance to a certain extent (45). MRI, an imaging technique with high resolution, can provide rich information on tissue structure and function, which can supplement the image information that cannot be provided by traditional clinical variables. Compared with a single clinical parameter, MRI images can intuitively display not only morphological features such as tumor size, location, and envelope invasion but also multiparameter imaging techniques (T2WI, DWI, DCE-MRI, etc.) that can reflect the biological behavior of tumors at the functional and metabolic levels, providing rich feature dimensions for prediction models. In particular, MRI-based imageomics technology can perform high-throughput, automated quantitative feature extraction for quantitative features and convert images into high-dimensional, mineable data. Through subsequent complex data analysis, researchers can obtain additional information beyond the original image and mine hundreds of quantitative features, such as tumor texture, morphology, and histogram. For example, SizeZoneNon-Uniformity can reflect the distribution variability of the volume of different sized regions in the image, a parameter that is closely related to tumor heterogeneity, and can assist in determining the aggressiveness of the tumor. ZoneEntropy can effectively differentiate the pathological grade of PCa by analyzing the complexity of the grayscale distribution of the image (30). These features extracted from MRI images provide richer data dimensions for the prediction model, which significantly improves the accuracy of BCR prediction; in addition, the Gleason score is an objective assessment tool for the severity of PCa malignancy, which is important for guiding clinical treatment and predicting prognosis (46). Higher Gleason scores indicate less differentiated, more malignant, and more invasive tumor cells (47), which predisposes patients to BCR. The assessment of these common factors should be emphasized in clinical work to identify risk groups efficiently. For high-risk patients, clinical staff should provide good preoperative education to improve patients’ knowledge of the risk of BCR and strengthen the follow-up management of patients after surgery. Notably, the pathophysiologic mechanism of BCR in patients after RP is still unclear, and most of the current studies on the risk factors for BCR are retrospective observational studies. To increase the reliability of the results, mechanism studies should be strengthened in the future, and large prospective studies should be conducted to observe the risk factor indicators several times at longer follow-up time points to improve the understanding of the potential risk factors by researchers and clinical staff.

Inspiration for future research

Predicting the risk of postoperative BCR after RP can help clinicians to accurately screen patients at high risk of postoperative recurrence and provide an evidence-based ground for strengthening the frequency of postoperative follow-up and formulating individualized adjuvant treatment plans. Overall, the prediction model based on multiparametric MR image features and other clinical variables significantly improved the prediction of postoperative BCR after RP. In this study, we found that research on BCR risk prediction modeling for patients after MRI examination of RP in China was carried out relatively late and is still in the development stage. First, the existing studies all used retrospective data analysis, with small sample sizes and insufficient diversity. In subsequent studies, higher-quality prediction models should be constructed in strict compliance with methodological guidelines, and multicenter and large-scale prospective studies should be vigorously carried out to confirm and optimize the validity and generalizability of the MRI prediction of BCR after RP. When applying the prediction model to clinical work, medical personnel should pay attention to combining the individual characteristics of high-risk groups, optimize the prediction model in a timely manner, and continuously calibrate the model. This can help clinical workers to provide appropriate interventions for high-risk groups, thus guaranteeing the patient’s outcome and at the same time alleviating the patient’s economic burden and reducing healthcare costs. Second, because PCa often presents as multifocal and heterogeneous, conventional MRI analysis methods are able to provide only subjective qualitative features and a few quantitative parameters, which makes it less useful in predicting the BCR after RP, particularly for small tumors and microscopic metastases (5). With the continuous development and advancement of imaging technology, the integration of imaging data such as positron emission CT and ultrasound elastography, into the prediction model can be considered a way to improve the accuracy of the model. Finally, with the accelerated generation of imaging data and the increasing volume of imaging data, the processing of MRI data and the extraction of useful information have increased requirements. With respect to information extraction, artificial intelligence (AI) can use complex algorithms to extract massive features that cannot be described intuitively by the naked eye and quantify tumor information (5). These features can be combined with patients’ clinical information to establish a prediction model, which can effectively solve the problem of tumor heterogeneity, which is difficult to assess quantitatively, and provide clinicians with a highly efficient and intelligent auxiliary decision support system. Future studies should exploit AI to lay the data foundation for establishing high-precision risk prediction models. In addition, the models can be transformed into technological tools such as online calculators and apps, fully enabling the positive promotion of AI and data mining technologies to the healthcare industry.

There are several limitations to this study: (I) only the Chinese and English literature on BCR risk prediction models for post-RP patients in China based on MRI examination were included in this study, which may have led to publication bias; (II) some articles did not report the specificity, sensitivity, or other indices, and the performance of the models in this study was evaluated only by the AUC or the C-index; and (III) the heterogeneity in aspects such as the research design, data sources, and modeling methods of the included literature. Only a qualitative analysis of the results was conducted, which may have led to limitations in the evaluation results.

Conclusions

A total of 17 studies involving 41 BCR risk prediction models for Chinese post-RP patients based on MRI examination were included in this study, and the model characteristics were analyzed descriptively. The results revealed that the predictive performance and applicability of the existing models were good, but the overall risk of bias was high. In the future, more high-quality, multicenter, large-sample studies should be conducted to construct and validate more practical BCR risk prediction models to provide clinicians with a high-quality basis for clinical decision-making.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.