Utilizing Deep Machine Learning for Prognostication of Oral Squamous Cell Carcinoma—A Systematic Review

The application of deep machine learning, a subfield of artificial intelligence, has become a growing area of interest in predictive medicine in recent years. The deep machine learning approach has been used to analyze imaging and radiomics and to develop models that have the potential to assist the clinicians to make an informed and guided decision that can assist to improve patient outcomes. Improved prognostication of oral squamous cell carcinoma (OSCC) will greatly benefit the clinical management of oral cancer patients. This review examines the recent development in the field of deep learning for OSCC prognostication. The search was carried out using five different databases—PubMed, Scopus, OvidMedline, Web of Science, and Institute of Electrical and Electronic Engineers (IEEE). The search was carried time from inception until 15 May 2021. There were 34 studies that have used deep machine learning for the prognostication of OSCC. The majority of these studies used a convolutional neural network (CNN). This review showed that a range of novel imaging modalities such as computed tomography (or enhanced computed tomography) images and spectra data have shown significant applicability to improve OSCC outcomes. The average specificity, sensitivity, area under receiving operating characteristics curve [AUC]), and accuracy for studies that used spectra data were 0.97, 0.99, 0.96, and 96.6%, respectively. Conversely, the corresponding average values for these parameters for computed tomography images were 0.84, 0.81, 0.967, and 81.8%, respectively. Ethical concerns such as privacy and confidentiality, data and model bias, peer disagreement, responsibility gap, patient-clinician relationship, and patient autonomy have limited the widespread adoption of these models in daily clinical practices. The accumulated evidence indicates that deep machine learning models have great potential in the prognostication of OSCC. This approach offers a more generic model that requires less data engineering with improved accuracy.

The application of deep machine learning, a subfield of artificial intelligence, has become a growing area of interest in predictive medicine in recent years. The deep machine learning approach has been used to analyze imaging and radiomics and to develop models that have the potential to assist the clinicians to make an informed and guided decision that can assist to improve patient outcomes. Improved prognostication of oral squamous cell carcinoma (OSCC) will greatly benefit the clinical management of oral cancer patients. This review examines the recent development in the field of deep learning for OSCC prognostication. The search was carried out using five different databases-PubMed, Scopus, OvidMedline, Web of Science, and Institute of Electrical and Electronic Engineers (IEEE). The search was carried time from inception until 15 May 2021. There were 34 studies that have used deep machine learning for the prognostication of OSCC. The majority of these studies used a convolutional neural network (CNN). This review showed that a range of novel imaging modalities such as computed tomography (or enhanced computed tomography) images and spectra data have shown significant applicability to improve OSCC outcomes. The average specificity, sensitivity, area under receiving operating characteristics curve [AUC]), and accuracy for studies that used spectra data were 0.97, 0.99, 0.96, and 96.6%, respectively. Conversely, the corresponding average values for these parameters for computed tomography images were 0.84, 0.81, 0.967, and 81.8%, respectively. Ethical concerns such as privacy and confidentiality, data and model bias, peer disagreement, responsibility gap, patient-clinician relationship, and patient autonomy have limited the widespread adoption of these models in daily clinical practices. The accumulated evidence indicates that deep machine learning models have great potential in the prognostication of OSCC. This approach offers a more generic model that requires less data engineering with improved accuracy.
Keywords: machine learning, deep learning, oral cancer, prognostication, systematic reveiw INTRODUCTION A total of 377, 713 new cases of oral cavity and lip cancer and 177, 757 deaths related to oral cancer were reported in the year 2020 [1]. Considering the location of oral squamous cell carcinoma (OSCC) and the corresponding aggressive behavior of this disease, it has been reported to have significant effects on the patients' post-treatment quality of life [2]. Recently, clear advances in diagnostic techniques and treatment modalities have been achieved [3]. However, OSCC is still characterized by a low average survival rate [4]. Accurate prognostication remains of utmost importance to improve survival rates [5].
Traditionally, the treatment of cancer depends mainly on tumor staging. However, staging discrepancies have contributed to inaccurate prognostication in OSCC patients [2]. Despite the increasing number of prognostic markers, the overall prognosis of the disease has not changed significantly [6]. This may be due to the challenges in the integration of these markers in the current staging system [7,8]. Additionally, individualized treatment of patient on a case-by-case basis is lacking. Therefore, improved diagnostic and prognostic accuracy could significantly assist the clinicians in making informed decisions regarding appropriate treatment for better survival [9].
To this end, machine learning techniques (shallow learning) have been reported to offer improved prognostication of OSCC [9,10]. Of note, the use of machine learning has been reported to provide a more accurate prognostication than the traditional statistical analyses [9,[11][12][13][14]. Machine learning techniques have been able to show promising results because they are able to discern the complex relationships between the variables contained in the dataset [9]. Considering the touted feasibility and benefits of the machine learning techniques in cancer prognostication, its application in this field has attracted significant attention in recent years. This is because it is poised to assist the clinicians in making informed decisions thereby improving and promoting better management of patient health. Interestingly, the advancements in technology have led to the modification of shallow machine learning to deep machine learning. This deep learning approach has also been touted to improve cancer management.
In this study, we aim to systematically review the published studies that have utilized deep machine learning techniques for OSCC prognostication. This is necessary to show the stateof-the-art performance of deep learning analytic methods for prognostication of the disease. Thus, the focused question was: "Does deep machine learning technique play a role in improving prognostication accuracy and guiding clinicians in making an informed decision."

Search Protocol
Detailed literature searches were performed using databases such as OvidMedline, PubMed, Scopus, Web of Science, and Institute of Electrical and Electronics Engineers (IEEE) from their inception until 15 May 2021. RefWorks software was used to properly manage the potentially relevant articles and remove any duplicate articles. Additionally, the reference lists of the included articles were manually searched to ensure that all the relevant articles have been included.

Search Strategy
The search approach was developed by combining search keywords: [(("oral cancer" OR "oral squamous cell carcinoma" OR "pre-cancerous" OR "oral potentially malignant") AND ("deep learning"))].

Inclusion Criteria
The Population, Exposure, Comparator, Outcomes, and Study design (PECOS) framework was used to define the research question(s) of this review. Thus, the P in the PECOS framework represents population (patients) with OSCC; E depicts that deep machine learning has been applied for prognostication, C ensures that the parameter of interest have OSCC patients with or without this parameter, O indicates that there is a clear outcome to be determined by the deep learning techniques, and S indicates that observational studies and/or clinical trials were also considered. Thus, original, observational, and clinical trials that utilized deep learning techniques for prognostication in OSCC were included. Additionally, only studies published in the English language were considered.

Exclusion Criteria
Studies in languages other than English and those that did not utilize deep learning for prognostication in oral cancer were excluded. Case reports, editorials, surveys, book chapter, comparative papers, symposium articles, conference articles, short communications, abstracts, opinions, perspectives, invited reviews, and letters to the editor were also excluded.

Study Selection
The study selection process was carried out in two distinct phases. Firstly, the titles and abstracts of potentially relevant articles were examined after the removal of duplicates. This phase was conducted by two independent reviewers (R.A., & O.Y.). A data extraction sheet was used for this process to ensure proper documentation with a Cohen's Kappa coefficient (κ = 0.91) for inter-observer reliability. This stage was followed by consensus meeting and discussion to resolve possible discrepancies before the study could be included in this review. For the second phase, these two independent reviewers extracted relevant information relating to the study characteristics of each of these potentially relevant articles.

Quality Assessment of the Included Studies
The Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) methodology was used to document the searching and screening processes in this study (Figure 1) [46,47]. The quality of the included studies were accessed using the Prediction model Risk of Bias Assessment Tool (PROBAST) as shown in Table 2. Additionally, PROBAST was used to evaluate and assess the risk of bias (ROB) of the potential studies to be included in this review. As the study examines deep machine learning method, the predictors' parameter from the PROBAST tool was modified to include the robustness of the methodology used in the included studies.

Results of the Database Search
A total of 34 studies met the eligibility criteria and were included in this review [2,48]. The details of the study selection process have been described using the PRISMA flowchart (Figure 1) [46]. The included studies that utilized deep learning for prognostication of OSCC are summarized in Table 1.
These studies concluded that the deep learning techniques could offer assistance to the clinicians in making informed decisions regarding choosing treatment options to avoid undertreatment or unnecessary treatments and thus achieving better management of the disease.
Considering the reported performance metrics (specificity, sensitivity, and accuracy) and the accumulated evidence presented in the included studies, deep machine learning models  + Indicates Low ROB/Low concern regarding applicability. − Indicates High ROB/high concern regarding applicability. ? Indicates unclear ROB/unclear concern regarding applicability.
have great potential in the prognostication of OSCC. This approach offers a more generic model that requires less data engineering with improved accuracy. A single study reported the performance of deep learning with four different performance metrics (sensitivity, specificity, accuracy, and area under receiving operating characteristics curve [AUC]) [16]. Similarly, a total of 11 studies reported the combination of the trio of sensitivity, specificity, and accuracy as the performance metrics for the deep machine learning method [15, 19-21, 24, 25, 30, 35, 37, 38, 42]. Both specificity and sensitivity were used to depict the performance of the model [17,20,22,27,48]. Additionally, specificity and accuracy were also used to demonstrate the performance of the deep learning model for prognostication in OSCC [23]. Other studies used either accuracy, C-index (concordance index), F1-score, or Dice similarity coefficient (Dsc) mean value as the performance metrics for reporting the potential benefits of the deep learning model [2, 18, 18, 26, 28, 29, 31-33, 39-41, 44, 45, 49].

Quality Assessment of the Studies Included in the Review
According to the PROBAST assessment, most (91.2%) of the included studies showed an overall low risk of bias and also exhibited low concern regarding applicability ( Table 2).

DISCUSSION
In this systematic review, the utilization of deep machine learning for prognostication in oral squamous cell carcinoma was examined. The deep learning methodology had been used to analyze various types of medical data such as clinicopathologic, histopathologic, gene expression, image, Raman spectroscopy, saliva metabolites, and computed tomography for better prognostication in OSCC. This review showed that a range of novel imaging modalities such as computed tomography (or enhanced computed tomography) images and spectra data have shown significant applicability to improve OSCC outcomes. Hence, deep machine learning methodology combined with medical imaging data can offer better and improved prognostication of OSCC. This can significantly assist the clinical management of patients with the disease [50].
The performance of the deep learning technique was mostly reported with either the combination of sensitivity, specificity, and accuracy or using a single performance metric. Based on the reported accuracy of the deep machine learning techniques in the included studies, it is evident that the deep machine learning technique can play a significant role toward the improved prognostication of oral cancer and guide clinicians in making informed decisions. The approach of using deep learning for prognostication can provide low-cost screening [19,36], smartphone-based solution [17,23], deep learning-based automatic prognostication [18,27,32], and early detection and prediction of outcomes [15,17,23,24,[37][38][39]42].
The afore-mentioned diagnostic ability and prognostication can greatly benefit the clinical management of OSCC patients [50]. For instance, deep learning can assist the pathologists in the effective multi-class grading, thereby, assisting in the timely and effective treatment protocol for the patients [32]. This can reduce operational workload and the possibility of burnout for the pathologists and enhance the proper management of the disease through timely grading [32]. Similarly, the deep learning model is capable of stratifying the patients into high-risk patients where they could be assigned to a more aggressive regimen or low-risk where more conservative treatment may be enough. This informed decision could assist in the overall survival of these patients by reducing the possibility of side effects such as hormonal disorder, trismus, or dental disease [52,53].
The availability of medical data in different formats (multiomics data-genomic, expression, proteomic, transcriptomic, and clinicopathologic data) through various databases such as the cancer genome atlas (TCGA), gene expression omnibus (GEO) has emerged as a great challenge to the traditional statistical methods of cancer prognostication [50]. Additionally, with the increase in computational power, advancement in technology (neural network model architecture), availability of medical dataset, the widely used shallow machine learning techniques have been modified to produce a deep machine learning technique, also known as the deep neural network (DNN) [54]. Interestingly, shallow machine learning has been reported to show promising results in various prognostication tasks such as prediction of locoregional recurrences [9,10], survival [55], occult nodal metastasis [56], and performed better than other methods such as nomograms [57].
Despite these promising results by the shallow machine learning techniques, the deep machine learning techniques have been reported to perform equally or outperform the shallow machine learning method [50,58,59] as it is more flexible, requires less feature engineering, and consist of complex layers and multiple neurons in each layer [50,60,61] (Figure 2). This gives deep machine learning a better predictive power [50]. An example of the deep neural network commonly used in cancer prognostication is the convolutional neural network (CNN) which is usually used for medical image data [50] (Figure 2). In CNN, the convolution and max-pooling layer are responsible for feature extraction of the input data [62] (Figure 2). While the convolution layers facilitate feature extraction from the image data, the pooling layers ensure that overfitting is minimized. The results from the convolution and pooling layers are passed to the fully connected layer for classification into labels (output) [62]. Apart from the CNN, the recurrent neural network (RNN) is another type of deep neural network which is suitable for text and sequence data [50]. In spite of the prospects of the deep learning models to improve OSCC outcomes through improved detection and diagnosis, most of these models have found widespread adoption in daily clinical practices. Several reasons have been attributed to the limited use of these models in clinical practice. A recent study showed that ethical concerns limited the potential use of these models in actual practices [63]. These ethical concerns include privacy and confidentiality, data and model bias, peer disagreement, responsibility gap, patient-clinician relationship, and patient autonomy [63]. Similarly, a recent study by Alabi et al., highlighted the concerns that are either inherent to the science of machine learning (technical) or the actual clinical implementation [64]. These include black box concern, amount of data, interpretability, explainability, and generalizability [64].
The strength of this systematic literature review is that it specifically examined the published studies that had examined deep learning in OSCC. This approach ensured that the contribution of the state-of-the-art deep learning techniques in OSCC was specifically examined. In addition, it offers the opportunity to understand the future research avenue of the application of deep learning in OSCC. An example of an exciting research area would be the development of new data fusion algorithms for improved prognostication in disease.
The main limitation is that most of the included studies used different performance metrics for the evaluation of the deep learning techniques. Similarly, the deep learning techniques used different data types in the analyses. Thus, it was challenging to make an insightful conclusion on the performance of these deep learning technique. Additionally, the dataset used to train the model was relatively small in most of the studies. Most of the developed deep learning models in the published studies were not externally validated. The study by Alhadi et al., provided an update on staging and World Health Organization grading as reliable OSCC prognostic indicators [65]. To the best of our knowledge, there is a dearth of published studies that have examined the application of machine learning for staging. Therefore, this serves as a potential area of further research in the future.
In conclusion, there is an increase in the application of deep learning for prognostication in OSCC. The deep learning models are poised to predict cancer prognosis more accurately. Thereby, offering precision and personalized management of the disease. It has shown to be better or equivalent to the current approaches in daily clinical practices. It is expected that the deep learning techniques can assist in the proper management of OSCC through improved diagnostic performance, insightful clinical decision making, streamline clinicians' work, offer a potential to reduce cancer care costs in the screening, and an effective assessment and surveillance of the disease. Thus, the clinicians and patients can spend more time in communication and in making shared decisions to improve the quality of care. In the future, it is important to develop deep learning models that combine multiple datasets from multiple modalities.

SUMMARY POINTS What Was Already Known on the Topic
There are several published studies on the application of machine learning techniques to analyze oral squamous cell carcinoma (OSCC).

What Knowledge This Study Adds
This study systematically reviewed the published studies that examined the application of deep machine learning techniques for prognostication in OSCC.
The majority of these studies used a convolutional neural network (CNN).
This review showed that a range of novel imaging modalities such as computed tomography (or enhanced computed tomography) images and spectra data have shown significant applicability to improve oral cancer outcomes.
The study concluded that the deep learning techniques could offer assistance to the clinicians in making informed decisions regarding choosing treatment options to avoid under-treatment or unnecessary treatments for the better management of OSCC.

DATA AVAILABILITY STATEMENT
The original contributions generated for the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author/s.