Gynecological cancer prognosis using machine learning techniques: A systematic review of the last three decades (1990–2022)
Introduction
The National Cancer Institute defines cancer as a “disease in which some of the body's cells grow uncontrollably and spread to other parts of the body” [1]. When such uncontrolled growth occurs in women's reproductive organs or genitals, they are referred to as ‘gynecological cancers’. There are five main types of gynecological cancers (cervical, ovarian, uterine, vaginal and vulval), named after the organ or tissue from which they originate. The most common gynecological malignancies – ovarian, cervical and uterine cancers – present a significant disease burden worldwide [2]. These cancers are prognostically variable. Of these, ovarian cancer has the highest rate of recurrence, at 85 % [3], and lowest rate of five-year survival at 30 % [4]. This is worsened by its non-specific symptomology and frequent late-stage diagnosis [5]. Although some clinical factors are prognostic, such as grade, stage, tumor subtype and debulking surgery success, the most commonly used clinical predictors [5] and biomarkers [6] are inadequate to predict clinical outcomes. Cervical cancer is one of the most common gynecological malignancies and is the fourth highest cause of cancer mortality in women worldwide [2]. Although some high Human Development Index (HDI) nations have had success in reducing the disease burden with screening and prevention programs [7], the prognostication of advanced-stage cervical cancer is variable [8]. Finally, although uterine cancer has a better prognosis than other malignancies, this disease is often very heterogeneous, making prognostication with current methods a challenge [9].
In recent years, prognostication has developed as a major focus in oncology, where decision making is influenced by the predicted probability of future events [10]. Treatment for gynecological cancers depends on the extent to which they have spread and the type of cancer, and includes modalities such as surgery, chemotherapy and radiotherapy. Developing oncological prediction algorithms and decision support tools would be useful for allowing clinicians to choose optimal screening, therapeutic and follow-up pathways for patients. However, challenges arise from the cancers' biological complexity and prognostic variability, alongside the ever-changing clinical, biological and pathological understanding of these malignancies [11].
Healthcare systems and clinicians currently use several tools to screen, diagnose and treat patients; however, current clinical approaches for many malignancies favor clinical staging and histopathological parameters with multivariate modelling showing limited success [12]. To address these shortcomings, machine learning (ML) approaches have been used to facilitate complex prognostic modelling that may outperform traditional methods [13]. ML methods aim to develop predictive algorithms without requiring complete prior rule definition, a valuable approach in complex clinical settings [12]. Predictive systems begin with data that undergoes pre-processing and feature extraction, followed by statistical analysis of extracted features, with selected features producing a classification result (Fig. 1). ML systems can be used at each step of this process. The ML classifiers (shown in Appendix A) used to perform the classification tasks include both unsupervised learning, which draws correlations within a dataset without a directed outcome, and supervised learning methods, including support vector machines (SVMs) and artificial neural networks (ANNs), which are goal-directed toward a particular outcome, regression or classification [12]. ML systems typically undergo training on a ‘training’ dataset and use a ‘validation’ dataset that the system is naïve to, to facilitate assessment of its performance, while still tuning its parameters. Finally, the system is typically exposed to a ‘testing’ dataset that it is naïve to, to facilitate an unbiased assessment of the final model's performance.
There is a paucity of clinical translatability of these methods for gynecological malignancies. Although they have been studied in this setting, they are variable in both approach and success. Systematic reviews have previously broadly summarized artificial intelligence (AI) in gynecologic imaging [14], or the application of ML methods broadly to gynecological cancers [15], [16]. However, to the authors' knowledge, a systematic review of the literature specific to prognostication in gynecological cancer has yet to be performed. Therefore, this study aims to systematically review ML in the prognostication of gynecological malignancies and evaluate the methodologies used.
Section snippets
Search strategy
This study was performed in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [17]. We searched PubMed, Embase, Web of Science, ENGINE, Scopus, IEEE Xplore and ACM Digital Library for studies exploring the use of ML methods for predicting prognosis for gynecological malignancies. Our search query was developed in PubMed using MeSH and keyword terms, and then revised for the other databases (Supplementary Table 1). The final iteration of
Results
The aim of this study was to systematically review the use of ML in the prognostication of gynecological malignancies and evaluate the methodologies used. In total, the initial search yielded 2207 unique papers with 349 papers passing title and abstract screening. 139 papers met all criteria for inclusion in the study [20], [21], [22], [23], [24], [25], [26], [27], [28], [29], [30], [31], [32], [33], [34], [35], [36], [37], [38], [39], [40], [41], [42], [43], [44], [45], [46], [47], [48], [49],
Discussion
This review aimed to systematically review and evaluate the methodologies used when applying ML to prognosticate common gynecological malignancies. The results show some promise in the field of ovarian, uterine and cervical cancers, and demonstrated discriminate predictive ability and superiority when individual studies compared these methods to non-ML methods. However, there were frequent methodological and reporting shortcomings that limit any conclusions that can be drawn in this review,
Conclusions
It has been shown that the literature features ML models that may improve patient outcomes in the future with discriminate benefits over current methods; however, concerns regarding ROB and applicability exist with the currently available literature. Genomic and clinicopathological predictor variables, in combination with RF and SVM ML methods, have been the most commonly applied tools to date. However, there is significant heterogeneity in the field, and in recent times, unique ML methods have
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
The authors would like to gratefully acknowledge the literature search contributions made by Ms. Jacky Cribb and Ms. Kaye Cumming, and analysis contributions from Ms. Daisy Eunji Cho, Ms. Yukei Oo and Ms. Yu Jin Cha.
CRediT authorship contribution statement
Conceptualization, JS., HR., RA., RG., XT., XZ., YL., and SKC.; methodology, JS., HR., RA., RG., XT., XZ., YL., and SKC.; formal analysis, JS., HR., HWL., and SKC.; data curation, JS., HR., and SKC.; writing—original draft preparation, JS., and HR.; writing—review and editing, JS., HR., RA., HWL., RG., XT., XZ., YL., TG., and SKC.; project administration, JS., HR., and SKC.; funding acquisition, SKC. All authors have read and agreed to the published version of the manuscript.
Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Some of the researchers who completed this work were employed under the Australian government's Rural Health Multidisciplinary Training Program and had full independence for this project.
References (169)
Ovarian cancer
Lancet
(2014)Cervical cancer
Lancet
(2019)Endometrial cancer
Lancet
(2016)- et al.
A review of statistical and machine learning methods for modeling cancer risk using structured clinical data
Artif Intell Med
(2018) A systematic review on the use of artificial intelligence in gynecologic imaging – background, state of the art, and future directions
Gynecol Oncol
(2022)Machine learning applications in gynecological cancer: a critical review
Crit Rev Oncol Hematol
(2022)- et al.
Artificial intelligence in gynecologic cancers: current status and future challenges – a systematic review
Artif Intell Med
(2021) - et al.
Neural network analysis to predict treatment outcome
Ann Oncol
(1993) Clinical data analysis with the use of artificial neural networks (ANN) and principal component analysis (PCA) of patients with endometrial carcinoma
RepPractOncolRadiother
(2005)- et al.
Cancer survival classification using integrated data sets and intermediate information
Artif Intell Med
(2014)