Comparing the Effectiveness of Artificial Intelligence Models in Predicting Ovarian Cancer Survival: A Systematic Review

ABSTRACT

Background

This systematic review investigates the use of machine learning (ML) algorithms in predicting survival outcomes for ovarian cancer (OC) patients. Key prognostic endpoints, including overall survival (OS), recurrence‐free survival (RFS), progression‐free survival (PFS), and treatment response prediction (TRP), are examined to evaluate the effectiveness of these algorithms and identify significant features that influence predictive accuracy.

Recent Findings

A thorough search of four major databases—PubMed, Scopus, Web of Science, and Cochrane—resulted in 2400 articles published within the last decade, with 32 studies meeting the inclusion criteria. Notably, most publications emerged after 2021. Commonly used algorithms for survival prediction included random forest, support vector machines, logistic regression, XGBoost, and various deep learning models. Evaluation metrics such as area under the curve (AUC) (18 studies), concordance index (C‐index) (11 studies), and accuracy (11 studies) were frequently employed. Age at diagnosis, tumor stage, CA‐125 levels, and treatment‐related factors were consistently highlighted as significant predictors, emphasizing their relevance in OC prognosis.

Conclusion

ML models demonstrate considerable potential for predicting OC survival outcomes; however, challenges persist regarding model accuracy and interpretability. Incorporating diverse data types—such as clinical, imaging, and molecular datasets—holds promise for enhancing predictive capabilities. Future advancements will depend on integrating heterogeneous data sources with multimodal ML approaches, which are crucial for improving prognostic precision in OC.

Abbreviations

ABDSN

attention‐based deep survival network

AUC

area under the curve

CLAM

clustering‐constrained‐attention multiple‐instance learning

COX‐PH

Cox proportional hazards

CPH

Cox proportional hazards

CT

computed tomographyn

DL

deep learning

DL‐CPH

deep learning Cox proportional hazards

DT

decision tree

EL

ensemble learning

GA‐XG

Boost Genetic Algorithm XGBoost

GB

gradient boosting

GEO

Gene Expression Omnibus

GNN

graph neural network

KNN

K‐nearest neighbors

LR

logistic regression

ML

machine learning

MLDP

machine learning–derived prognostic signature

MRI

magnetic resonance imaging

NPV

negative predictive value

OS

overall survival

PFS

progression‐free survival

PLS

partial least squares

PPV

positive predictive value

PRISMA

Preferred Reporting Items for Systematic Reviews and Meta‐Analyses

RF

random forest

RFS

recurrence‐free survival

SEER

Surveillance, Epidemiology, and End Results

SMC

Samsung Medical Center

SVM

support vector machines

TCGA

The Cancer Genome Atlas

TRP

treatment response prediction

ViT‐DL

vision transformer–based deep learning

WSI

whole slide image

1. Introduction

Ovarian cancer (OC) is a significant global health concern, ranking seventh in prevalence and eighth in mortality among women worldwide [1]. It is the third leading cause of death among gynecological cancers, following uterine and cervical cancers, with over 300 000 cases diagnosed annually and nearly 200 000 deaths. Despite advancements in diagnosis and treatment, OC remains a formidable challenge due to its often asymptomatic nature in early stages, leading to late‐stage diagnoses and poor prognosis [2]. The global incidence of OC is estimated at 6.6 per 100 000 women per year, and it has been the fifth most common cause of cancer‐related death in women in the Western world over the last two decades [3].

OC tumors are classified into three categories as follows: benign, borderline, and malignant [4]. While malignant neoplasms often necessitate invasive surgical interventions, benign masses can be managed with safer monitoring approaches. The poor prognosis associated with OC is largely due to the advanced metastatic stages at which many patients are diagnosed [5]. Variations in OC incidence and mortality rates across different regions emphasize the need to understand the diverse epidemiology influenced by factors such as age, reproductive history, hormonal imbalances, genetic predispositions, and lifestyle choices [6]. A family history of OC, particularly involving first‐degree relatives, significantly increases an individual's risk, often due to hereditary mutations in the BRCA1 and BRCA2 genes [7].

Despite substantial progress in treatment options, the overall survival (OS) rate for OC patients has seen only modest improvements. The GLOBOCAN project estimates that by 2050, the global incidence of OC will rise by 55%, with the highest burden expected to fall on low‐ and middle‐income countries (LMICs) [2]. Approximately 75% of cases occur in postmenopausal individuals, with an incidence rate of 40 per 100 000 annually in those aged 50 and above. Early detection significantly boosts the 5‐year survival rate from 3% at Stage IV to 90% at Stage I [8]. Key factors influencing survival include cancer stage, tumor size, residual tumor after surgery, tissue type, and race [9]. The limited improvement in survival rates is partly attributable to the lack of effective prognostic biomarkers and screening methods, leading to delayed diagnoses [6].

Accurate prognosis and survival predictions are critical in managing OC, aiding in treatment decisions and reducing patient anxiety [10]. However, oncologists often overestimate survival, leading to inadequate end‐of‐life care. Studies, such as those by Alexi et al., have shown that oncologists' prognoses can be overly optimistic, especially in long‐term patient relationships, as is common in OC cases [11]. This highlights the need for reliable prognostic tools that can assist in clinical decision‐making, particularly in identifying patients nearing the end of life.

Traditional statistical methods have been widely used to predict cancer prognosis, including in OC [12, 13, 14]. For instance, the multivariable Cox proportional hazards (CPH) model has been employed to assess OS in OC patients [13]. However, the emergence of machine learning (ML) algorithms offers a promising alternative, providing enhanced accuracy and the ability to handle complex, nonlinear data patterns [15, 16]. ML, a subset of artificial intelligence, has demonstrated significant potential in survival analysis, particularly when combined with deep learning (DL) techniques. DL models, known for their layered architecture and capacity for automatic feature selection, have shown remarkable performance in various cancer survival predictions, including breast [17], cervical [18], lung [19], and bladder cancers [20].

Although ML and DL have shown advantages over traditional statistical models in various survival analyses, their potential for enhancing OC survival prediction remains to be fully explored [21]. This review aims to systematically analyze all original studies that have utilized ML algorithms to predict OC survival. The focus will be on evaluating models based on OS, recurrence‐free survival (RFS), progression‐free survival (PFS), and treatment response prediction (TRP). The review will assess each step of the modeling process, from data collection to performance evaluation, and will identify key variables influencing survival predictions, categorizing them based on survival type.

2. Methods

This systematic review was meticulously conducted in alignment with the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) guidelines, ensuring a comprehensive and standardized approach to data collection and analysis [22].

2.1. Information Sources

A literature search was systematically conducted across multiple databases, including PubMed, Scopus, Web of Science, and Cochrane Library. Additionally, reference lists of relevant studies were examined to ensure comprehensive coverage of the literature. The final search across all sources was completed in December 2024.

2.2. Search Strategy and Selection Process

The search strategy aimed to identify relevant studies in three primary domains: OC, ML, and survival prediction. A combination of controlled vocabulary (MESH terms) and keywords was employed to construct the search queries, details of which are provided in Supplementary file. A 10‐year date filter was applied to focus on the most recent research. The selection process involved an independent screening of titles and abstracts, and full texts by two authors to assess their eligibility. Discrepancies were addressed through consultation with a corresponding author, ensuring consistency and reducing bias. No automation tools were utilized during the selection process.

2.3. Eligibility Criteria

Inclusion and exclusion criteria were established prior to the review. Studies were included if they were original research articles published in English, employed ML models to predict OC survival, clearly presented the model's output and evaluation, and specifically focused on OC. Exclusion criteria encompassed books, letters to the editor, review articles, and meta‐analyses, as well as studies that did not focus on OC or lacked detailed information regarding input datasets, modeling processes, and model evaluation (Table 1).

TABLE 1.

Inclusion and exclusion criteria.

Inclusion criteria	Exclusion criteria
Only journals and original papers are used	Studies written in other languages
Studies written in English only	Other sources, such as books, theses, editorials, and systematic reviews
Papers that focused on ML– or DL–based ovarian cancer survival prediction	Papers that focused on statistical techniques for ovarian cancer survival prediction
	Studies without enough modeling details

2.4. Data Collection Process

Data extraction was conducted using a custom‐designed form tailored for this review. Two authors independently extracted relevant information from each study, including author details, publication year, country of origin, dataset characteristics, data types, ML models used, criteria for model performance assessment, and significant features influencing model outcomes. Data extraction focused on key outcomes, such as OS, PFS, RFS, and TRP. For studies reporting multiple outcomes or models, the data related to survival outcomes and the models demonstrating the best performance were prioritized (Table 2). Data synthesis involved summarizing the results through descriptive statistics and narrative synthesis.

TABLE 2.

Characteristics of included articles.

Author, country, year	Year	Dataset	Participants/data records	Survival outcome	Best ML model	Results (eval)
Chih‐Jen T, Taiwan, (2017) [23]	2017	Chung Shan Medical University Hospital Tumor Registry	987	RFS	C5.0	Accuracy (0.90)
Paik ES, Korea, (2019) [24]	2019	Samsung Medical Center, Asan Medical Center	1128	OS	GB	AUC (0.84)
Wang S, China, (2019) [25]	2019	Two hospitals	8917	RFS	DL‐CPH	C‐index (0.71)
Arezzo F, Italy, (2020) [26]	2020	Tertiary center (2018–2019)	64	PFS	RF	Accuracy (0.93)
						Sensitivity (0.90)
						Precision (0.90)
						AUC (0.92)
Kaur I, India, (2021) [27]	2021	Rajiv Gandhi Cancer Institute & Research Center	140	OS	EL	Accuracy (0.71)
						Sensitivity (0.79)
						Specificity (0.61)
						AUC (0.80)
Hsiao YW,Taiwan, (2021) [28]	2021	GEO	106	OS	GA‐XGBoost	Accuracy (0.58)
						Sensitivity (0.82)
						Specificity (0.38)
						F1‐score (0.65)
He T, India, (2021) [29]	2021	TCGA GEO	790	OS	RF‐LR‐COX	C‐index (0.76)
Li HM, China, (2021) [30]	2021	Fudan University Shanghai Cancer Center (2014–2019)	117	RFS	DL—SVM	C‐index (0.62)
						AUC (0.85)
Shannon NB, Singapore, (2021) [31]	2021	GDSC, TCGA,GEO	—	TRP	LR XGB	Accuracy (0.82)
Feng Y, China, (2022) [32]	2022	Hospital (2009–2018)	98	OS	DT	AUC (0.69)
Sorayaie Azar A, Iran, (2022) [33]	2022	SEER	42 827	OS	RF	Accuracy (0.88)
						Sensitivity (0.71)
						Specificity (0.92)
						F1‐score (0.71)
						AUC (0.82)
Li Y, China, (2022) [34]	2022	TCGA,GEO	532	OS	XGBoost	AUC (0.72)
Zheng Y, China, (2022) [35]	2022	Qilu Hospital of Shandong University	734	OS	ViT‐DL	C‐index (0.74)
						AUC (0.82)
Chris S, USA, (2022) [36]	2022	Academic Cancer Institution	245	OS	EL	Accuracy (0.79)
						Sensitivity (0.71)
						Specificity(0.80)
						AUC (0.76)
Nero C, Italy, (2022) [37]	2022	Fondazione Policlinico Universitario (2016–2020)	664	PFS	DL	AUC (0.71)
						NPV (0.69)
						PPV (0.75)
Avesani G, Italy, (2022) [38]	2022	Multicentric database from four referral centers	218	PFS	XGBoost	AUC (0.62)
Tingyuan L, China, (2023) [39]	2023	TCGA	284	OS	DNN	AUC (0.94)
Meixuan W, China, (2023) [40]	2023	TCGA‐OV	90	OS	ABDSN	C‐index (0.58)
Qing H, China, (2023) [41]	2023	12 multicenter cohorts	2626	OS	MLDPS	C‐index (0.79)
Xiangmei L, China, (2023) [42]	2023	SEER (2004–2015)	1131	OS	LR	Accuracy (0.80)
						Precision (0.56)
						AUC (0.80)
Sheng W, China, (2023) [43]	2023	TCIA, TCGA,	489	OS	LASSO Regression	AUC (0.72)
David M, USA, (2023) [44]	2023	Academic institution	71	RFS	Elastic Net	C‐index (0.72)
Lili L, China, (2023) [45]	2023	Hospital of Chongqing Medical University (2013–2019)	3839	RFS	DL	C‐index (0.75)
						AUC (0.98)
Pradeep G, India, (2024) [46]	2024	OCD, SEER	9450	OS	Temporal GNN	Accuracy (0.95)
Lindong J, USA, (2024) [47]	2024	BRCA TCGA	355	OS	DL	C‐index (0.73)
Mirielle M, USA, (2024) [48]	2024	TCGA	360	OS	LR	Accuracy (0.64)
						Precision (0.68)
						Recall (0.34)
Yongmei H, USA, (2024) [49]	2024	SEER(2010–2015)	5656	OS	PLS Regression	C‐index (0.75)
Lian J, China, (2024) [50]	2024	Hospital (2013–2015)	102	PFS	RF	AUC (0.77)
Li‐Rong Y, China, (2024) [51]	2024	Yunnan Cancer Hospital (2012–2022)	1392	RFS	XGBoost	AUC (0.78)
						Sensitivity (0.73)
						Specificity (0.71)
						Accuracy (0.80)
Wan‐Chun L, Taiwan, (2024) [52]	2024	Two tertiary centers (2010–2019)	723	TRP	CatBoost	AUC (0.87)
						F1‐score (0.69)
Zijian Y, China, (2024) [53]	2024	TCGA‐OV	634	TRP	GNN
Qiwang L, China, (2024) [54]	2024	TCGA, GEO	381	TRP	RSF	C‐index (0.75)

3. Results

The initial search across the four specified databases yielded a total of 2400 articles. After applying filters to limit studies to those published within the last 10 years, excluding non‐original papers, and removing duplicates, 1172 articles remained. Subsequent screening resulted in the selection of 32 original articles that met the inclusion criteria for this systematic review (Figure 1).

FIGURE 1.

PRISMA diagram.

3.1. Study Characteristics

The characteristics of the studies included in this review are summarized in Table 2. Most studies were published post‐2021, with a significant concentration of research originating from Asia (24 articles), followed by the United States (5 articles), and Europe (3 articles). OS was the most frequently predicted outcome, featured in 18 studies, whereas RFS was addressed in 16 studies. PFS and TRP were each the focus of four studies. A diverse range of ML and DL models was employed for survival prediction across the studies (Table 3).

TABLE 3.

Categorizing the characteristics of the included studies.

Characteristics	Categories	OS (n)	PFS (n)	RFS (n)	TRP (n)
Geographic	Asia	14 (23, 27, 30–33, 44, 45, 47, 49–51, 71, 72)	1 (39)	5 (25, 28, 29, 40, 42)	4 (43, 48, 52, 53)
	Europe	—	3 (24, 37, 38)	—	—
	USA	4 (26, 34, 54, 55)	—	1 (41)	—
Public datasets	TCGA	7 (45, 47, 49, 51, 54, 55, 72)	—	—	3 (48, 52, 53)
	TCIA	1 (45)	—	—	—
	GDSC		—	—	1 (52)
	GEO	3 (44, 47, 49)	—	—	2 (52, 53)
	SEER	4 (23, 26, 30, 32)	—	—	—
	SMC	1 (27)	—	—	—
	AMC	1 (27)	—	—	—
Data types	Clinical	12 (23, 26, 27, 30, 32, 33, 45, 47, 50, 54, 71, 72)	3 (24, 37, 39)	6 (25, 28, 29, 40–42)	3 (43, 48, 53)
	WSI	1 (51)	1 (38)	—	1 (48)
	Ultrasound	—	1 (24)	—	—
	CT	2 (45, 50)	2 (37, 39)	1 (28)	1 (43)
	MRI	—	—	2 (25, 40)	—
	Molecular	7 (31, 45, 47, 49, 51, 54, 55)	1 (38)	—	1 (53)
Models	RF	5 (23, 30, 47, 71)	3 (24, 37, 39)	2 (29, 42)	3 (43, 52, 53)
	DT	4 (23, 30, 33)	—	1 (29)	1 (52)
	KNN	2 (23)	1 (24)	1 (29)	2 (43, 52)
	LR	7 (26, 30, 44, 45, 47, 55, 71)	2 (24, 37)	1 (29)	2 (43, 52)
	SVM	2 (23)	1 (37)	3 (29, 40, 42)	2 (43, 52)
	XGBoost	5 (23, 44, 47, 71)	1 (37)	1 (29)	2 (43, 52)
	AdaBoost	2 (23)	—	—	—
	DL	5 (32, 50, 51, 54, 72)	1 (38)	3 (25, 28, 40)	1 (48)
	Others	7 (27, 30, 31, 34, 49, 71)	—	2 (41, 42)	2 (43, 52)
Evaluations	Accuracy	7 (23, 30, 32, 34, 44, 55, 71)	1 (24)	2 (29, 42)	1 (52)
	Sensitivity	5 (23, 34, 44, 55, 71)	1 (24)	1 (29)	—
	Specificity	4 (23, 34, 44, 71)	—	1 (29)	—
	Precision	2 (30, 55)	1 (24)	—	—
	AUC	10 (23, 27, 30, 33, 34, 45, 47, 50, 71, 72)	4 (24, 37–39)	3 (25, 29, 40)	1 (43)
	C‐index	6 (26, 31, 49–51, 54)	—	4 (25, 28, 40, 41)	1 (53)
	Others	2 (23, 44)	2 (24, 38)	—	1 (43)

3.2. Datasets Characteristics

The datasets utilized in the reviewed studies were categorized into open‐ and closed‐access datasets. Open‐access datasets were utilized in 16 studies, all focusing on OS outcomes. Among these, 10 studies relied on datasets from The Cancer Genome Atlas (TCGA), five from the Gene Expression Omnibus (GEO), and four from the Surveillance, Epidemiology, and End Results (SEER) datasets. The SEER database provided the largest dataset, containing 42 827 records [33], whereas the smallest dataset included clinical data and ultrasound images for 64 patients [26]. Nine studies utilized datasets containing over 1000 samples [24, 25, 33, 41, 42, 45, 46, 49, 51], with six studies concentrating on OS using tabular data [24, 33, 41, 42, 46, 49] and two focusing on RFS using imaging data [25, 45]. Fifteen studies utilized closed‐access datasets sourced from healthcare institutions. Among these, five studies focused on OS [27, 32, 35, 36, 41], four on PFS [26, 37, 38, 50], six on RFS [23, 25, 30, 44, 45, 51], and one on TRP [52].

3.3. Data Cleaning and Preprocessing

Data wrangling was performed across all studies, though detailed descriptions of these processes were often sparse. Common data cleaning steps included the removal of duplicate records, correction of structural errors, addressing missing data, and identification of outliers [27, 28, 33]. Statistical methods and ML models were commonly utilized to manage missing data. For instance, some studies employed imputation methods using the mean, median, or mode [27], whereas others applied ML models such as K‐nearest neighbor (KNN) [24, 27], Elastic Net algorithm [43], multiple chained equations [36], and random forest (RF) [50]. In one study, features with over 50% missing data were excluded from analyses [27]. Feature selection was predominantly carried out using supervised and unsupervised ML algorithms, including Fisher's exact test [55], recursive feature elimination (RFE) [26], LASSO [30, 34, 43, 50], convolutional neural networks (CNNs) [25, 37, 38, 45], RF [34], graph neural network (GNN) [53], and XGBoost [26, 28, 34]. Prior to modeling, the distribution and independence of features were assessed using various statistical tests, including T‐tests, Wilcoxon, Mann–Whitney, McNemar, the chi‐squared test, Pearson, and Spearman correlations [24, 26, 29, 33, 34, 35]. One study [32] utilized the National Comprehensive Cancer Network (NCCN) guidelines to inform feature selection.

3.4. Modeling

The most frequently used models for survival prediction were RF, support vector machine (SVM), logistic regression (LR), XGBoost, and DL algorithms. RF and XGBoost were the predominant models applied to OS prediction, whereas RF and LR were common in PFS prediction, and DL was primarily used in RFS prediction. Although most included studies employed multiple ML models for OC survival prediction, some studies opted for simpler ML models, largely due to the reliance on traditional statistical methods for prognosis.

Innovative ML models were explored in several studies for OC prognosis, including vision transformers (ViTs) [35], bagging‐based models [28], attention‐based learning [37, 40], graph‐based learning [53], and fusion models [41, 45, 46]. Among studies comparing multiple models, XGBoost achieved the best performance in four studies [28, 31, 38, 51], followed by RF [26, 33, 54]. Other models, such as bagging [27], gradient boosting [24], CatBoost [52], and LR [31, 42], each demonstrated superior performance in individual studies. Cross‐validation was the most commonly used method for training and validating ML models [26, 27, 28, 29, 31, 32, 37, 38, 50], whereas a few studies employed leave‐one‐out validation [30, 55]. In some studies [25, 26, 30, 38, 45, 53], DL models were used as feature extractors, with ML algorithms subsequently applied for survival prediction.

3.5. Model Optimization and Evaluation

Meta‐parameter tuning was employed in some studies to optimize ML models, using methods such as grid search, random search, and Bayesian optimization [36, 37, 47]. Regularization, batch normalization, and learning rate scheduling were also applied in several studies to optimize DL models [25, 37, 38]. Evaluation methods varied significantly, with nine studies implementing external validation [24, 25, 28, 29, 34, 38, 52, 53, 54], whereas the remaining studies relied on internal validation. Among evaluation metrics, the area under the curve (AUC) was the most frequently reported, appearing in 18 studies, followed by accuracy in 11 studies, sensitivity in seven studies, and the C‐index in 11 studies (Table 3). Studies often allocated between 10% and 30% of their data for ML model evaluation [31, 32, 43, 45, 47]. The maximum and minimum outputs of ML models based on each evaluation criterion were compiled (Table 4).

TABLE 4.

Classification of evaluation criteria used based on the type of survival (from the lowest to the highest).

Evaluation criteria	OS		PFS		RFS		TRP
	Min	Max	Min	Max	Min	Max	Min	Max
AUC	0.64 (55)	0.94 (72)	0.62 (37)	0.92 (24)	0.78 (29)	0.98 (25)	0.87 (43)	0.87 (43)
Accuracy	0.58 (44)	0.95 (32)	0.93 (24)	0.93 (24)	0.80 (29)	0.90 (42)	0.82 (52)	0.82 (52)
Sensitivity	0.34 (55)	0.82 (44)	0.90 (24)	0.90 (24)	0.73 (29)	0.73 (29)	—	—
Specificity	0.38 (44)	0.92 (23)	—	—	0.71 (29)	0.71 (29)	—	—
C‐index	0.73 (54)	0.79 (31)	—	—	0.62 (40)	0.75 (25)	0.75 (53)	0.75 (53)

3.6. Influential Variables

Most of the included studies identified key clinical variables that significantly influenced the prediction of OC survival. Among these, age at diagnosis, tumor stage, histology, tumor grade, chemotherapy, metastasis status, race, CA‐125 levels, lymph node involvement, surgery information, and tumor diameters consistently emerged as critical determinants. Six studies also integrated molecular data, identifying factors such as BRCA1, NBN, BRIP1, RAD50, PTEN, and PMS2, which positively correlated with survival, whereas FANCE, FOXM1, KRAS, FANCD2, TTN, and CSMD3 were associated with poorer outcomes [48]. Additional influential variables included CCR5 expression levels [43], CT–based radiomics [43], and serum proteomics [44]. Furthermore, health‐related quality of life and psychosocial factors were highlighted in several studies as important contributors to survival outcomes [36]. Collectively, these variables provide a comprehensive framework for improving prognostic accuracy in OC (Table 5).

TABLE 5.

Significant variables in ovarian cancer survival prediction.

Important variables	OS	PFS	RFS	TRP
Age at diagnosis	9 (23, 26, 27, 30, 33, 50, 51, 54, 72)	1 (24)	1 (42)	—
Tumor stage	5 (23, 26, 27, 50, 54)	1 (39)	2 (28, 42)	—
Histology	3 (23, 26, 27)	—	—	—
Chemotherapy	3 (23, 30, 71)	1 (24)	—	—
Tumor grade	4 (23, 26, 27, 51)	—	—	—
Metastasis	—	1 (24)	—	—
Race	3 (23, 26, 54)	—	—	—
CA‐125 levels	3 (27, 33, 50)	2 (39, 41)	2 (28, 41)	—
Lymph node status	—	1 (24)	—	—
Surgery information	1 (27)	—	—	—
Tumor diameters	2 (45, 50)	—	—	—
Others	11 (23, 27, 30, 31, 33, 34, 45, 54, 55, 71, 72)	2 (24, 39)	—	3 (43, 52, 53)

4. Discussion

This systematic review comprehensively analyzed 2400 articles across four databases, leading to the final inclusion of 32 original studies. These studies addressed OC prognosis, utilizing ML and DL models as predictive tools. The datasets utilized in these studies ranged significantly in size and type, from small collections to large‐scale datasets, and included tabular, image‐based, and molecular data. Key clinical variables frequently identified as influential in these studies include age at diagnosis, tumor stage, histology, and treatment types.

While AI models demonstrate promising accuracy compared to non‐AI methods for survival prediction, this potential can be misleading without rigorous validation and use of heterogeneous, real‐world datasets. The reliance on homogeneous, single‐institution datasets in some studies limits generalizability, highlighting the challenges of translating AI–driven detailed predictions into clinical practice [56].

Compared to traditional statistical models, interpretable ML models offer enhanced transparency and understandability by revealing feature influence on predictions, fostering clinical trust [57]. Furthermore, ML's ability to handle complex, large datasets and focus on individual patient characteristics, rather than population averages, enables more personalized survival predictions, highlighting their potential clinical utility [58].

Dataset selection is crucial for predictive model reliability and applicability; open‐access datasets, with larger samples and accessibility, enhance transparency, reproducibility, and collaboration. Conversely, small datasets pose overfitting risks, compromising generalizability, thus emphasizing careful dataset selection for advancing knowledge and scientific discovery [59, 60].

Data cleaning and preprocessing are fundamental for ensuring data quality, reducing errors, and mitigating bias, which are crucial for reliable predictive model training. Their methodological rigor directly impacts the validity and accuracy of survival prediction models [61, 62].

Imbalanced datasets, a frequent issue in medical research, are often addressed with techniques like SMOTE [32, 33] and cost‐sensitive [26] methods to reduce bias and enhance reliability. However, recent findings suggest these corrections may not improve predictive performance and can produce inaccurate probability estimates, highlighting the need for caution when applying such methods [63].

XGBoost and RF emerged as top‐performing ML algorithms in accuracy and interpretability, whereas decision trees remain favored in clinical settings for their simplicity and ease of implementation [64].

The integration of clinical, imaging, and molecular data significantly enhances the accuracy and reliability of ML models for OC survival prediction, with multidimensional approaches showing superior predictive performance. Image datasets, supported by advanced processing algorithms, often outperform tabular data, with CNNs traditionally used for medical image analysis despite their focus on local information [12, 65]. Transformer models, utilizing self‐attention mechanisms, overcome this limitation by aggregating global information and capturing broader spatial patterns [35]. Additionally, GNNs effectively model relational data, such as molecular networks and tissue interactions, whereas ViTs excel in analyzing imaging data [66]. Integrating these approaches presents a potentially valuable strategy for constructing comprehensive models capable of generating dependable predictive assessments.

In OC survival studies, feature selection and model interpretability are crucial for developing clinically relevant ML models. Identifying key features like age and tumor stage improves model accuracy. Techniques such as SHAP [67] and LIME [68] enhance transparency by explaining feature influence, fostering trust and facilitating integration into personalized medicine for improved clinical decisions.

A significant limitation in the current body of literature is the lack of reporting on study quality and the tendency to overlook potential biases. For instance, achieving an accuracy rate above 90% might superficially suggest that prognostication challenges have been resolved. However, this assumption is misleading due to persistent issues like overfitting, insufficient transparency and interpretability in model development, and a lack of robust external validation. AI research in OC often does not sufficiently address these critical limitations, which are vital for accurately assessing the true performance and clinical applicability of these models.

Advancing the field requires addressing several key areas to bring AI–based models closer to clinical utility. First and foremost, the implementation of rigorous validation protocols must become standard practice. External validation employing diverse, multi‐institutional datasets is crucial. These methods are critical for evaluating the robustness of models and their ability to generalize to new, unseen data. Additionally, reporting confidence intervals and employing statistical tests for model comparison will provide clinicians with more reliable insights into the relative performance of these models, enabling more informed decision‐making in clinical settings.

Another crucial consideration is the integration of AI models into the clinical workflow. These models should not be perceived as standalone solutions but rather as tools designed to assist clinicians, particularly in navigating complex decision‐making processes. Future research should prioritize understanding how these models can function within clinical environments, possibly serving as assistive tools that offer second opinions or illuminate novel insights that may be missed by human clinicians. This approach emphasizes the role of AI as a complementary resource rather than a replacement for clinical expertise [69].

Finally, enhancing the transparency and reproducibility of AI research is imperative. Researchers must adhere to standardized reporting guidelines, such as the TRIPOD checklist [70], to ensure that all relevant aspects of their models and datasets are thoroughly documented. Moreover, making code and datasets openly accessible, whenever possible, would significantly improve the reproducibility of research findings and allow for more rigorous testing of model generalizability. The current scarcity of openly accessible, heterogeneous datasets in OC research represents a major obstacle that must be overcome to facilitate more meaningful and clinically translatable AI advancements.

5. Conclusion

OC is a highly heterogeneous disease characterized by various histologic subtypes, making accurate prognosis particularly challenging. The application of ML has shown promise in addressing these complexities, particularly in managing missing data and integrating multiple predictive algorithms. This study identified RF, SVM, LR, XGBoost, and DL as the most frequently utilized models for survival prediction in OC.

Key predictive factors consistently identified across models included age at diagnosis, tumor stage, histologic subtype, treatment type, and specific biomarkers. The integration of diverse datasets—encompassing clinical, imaging, and molecular data—was found to enhance the accuracy of survival predictions. This suggests that combining heterogeneous, multidimensional data with advanced ML techniques offers a more robust approach to predicting OC survival. Moving forward, the emphasis should be on refining these models and exploring their potential for broader clinical application.

Author Contributions

Farkhondeh Asadi: conceptualization, data curation, formal analysis, investigation, methodology, project administration, resources, supervision, validation, writing – original draft, writing – review and editing. Milad Rahimi: data curation, formal analysis, methodology, writing – original draft, writing – review and editing. Nahid Ramezanghorbani: writing – original draft, writing – review and editing. Sohrab Almasi: writing – original draft, writing – review and editing.

Ethics Statement

This study was approved by Shahid Beheshti University of Medical Sciences (IR.SBMU.RETECH.REC.1401.660).

Consent

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1. Supporting Information.

Data Availability Statement

All data used in the publication of this work were obtained from published studies. The abstracts for these studies are available in the PubMed, Scopus, Web of Science, and Cochrane database.