Effectiveness of Artificial Intelligence for Personalized Medicine in Neoplasms: A Systematic Review

Purpose Artificial intelligence (AI) techniques are used in precision medicine to explore novel genotypes and phenotypes data. The main aims of precision medicine include early diagnosis, screening, and personalized treatment regime for a patient based on genetic-oriented features and characteristics. The main objective of this study was to review AI techniques and their effectiveness in neoplasm precision medicine. Materials and Methods A comprehensive search was performed in Medline (through PubMed), Scopus, ISI Web of Science, IEEE Xplore, Embase, and Cochrane databases from inception to December 29, 2021, in order to identify the studies that used AI methods for cancer precision medicine and evaluate outcomes of the models. Results Sixty-three studies were included in this systematic review. The main AI approaches in 17 papers (26.9%) were linear and nonlinear categories (random forest or decision trees), and in 21 citations, rule-based systems and deep learning models were used. Notably, 62% of the articles were done in the United States and China. R package was the most frequent software, and breast and lung cancer were the most selected neoplasms in the papers. Out of 63 papers, in 34 articles, genomic data like gene expression, somatic mutation data, phenotype data, and proteomics with drug-response which is functional data was used as input in AI methods; in 16 papers' (25.3%) drug response, functional data was utilized in personalization of treatment. The maximum values of the assessment indicators such as accuracy, sensitivity, specificity, precision, recall, and area under the curve (AUC) in included studies were 0.99, 1.00, 0.96, 0.98, 0.99, and 0.9929, respectively. Conclusion The findings showed that in many cases, the use of artificial intelligence methods had effective application in personalized medicine.


Introduction
Cancer refers to a set of diseases in which some body's cells decide to divide continuously, and as a result, they spread into surrounding tissues ("National Cancer Institute"). Cancer is a genetic disease that changes genes' function and can control the way cells divide [1,2]. In 2018, there were 24.5 million cancer cases (16.8 million without nonmelanoma skin cancer [NMSC]) and 9.6 million cancer deaths worldwide. Most of the disability-adjusted life-years (DALYs) caused by cancer lead to 97% of lives lost and only 3% of living with a disability. The leading cause of cancer deaths and DALYs for men is related to TBL (Tracheal, Bronchus, and Lung) cancer (1.3 million deaths and 28.4 million DALYs). However, the common leading cause of cancer death and DALYs for women is related to breast cancer (601.000 deaths and 17.4 million DALYs) [3]. Due to the growing number of cancer cases globally, timely detection and selection of the best treatment are considered key steps. Detection of cancer in the early stages can significantly increase the possibilities of successful treatment [4]. Early detection of cancer is greatly influenced by two factors of early diagnosis and screening [5]. In recent years, due to the importance of analyzing the genetic profile of people with cancer, the method of using extensive genomic data in a new field called precision medicine has been introduced [6]. With precision medicine and the progression of nextgeneration sequence (NGS), patients' genomic profiles can be used for disease diagnosis, risk prediction, and treatment of diseases [7]. Thus, the main aims of precision medicine include early diagnosis, screening, and personalized treatment regime for patient based on genetic-oriented features and characteristics [8].
Precision medicine for treating diseases considers various factors, which can be referred to the genome of individuals, lifestyle, environmental factors, and characteristics of patients [9]. Precision medicine allows clinicians to select more effective and accurate therapeutic and preventive approaches to a specific illness such as cancer. It can work in subgroups of patients based on their genetic make-up, environmental factors, and lifestyle [10]. In cancer genomics, the multiomics data, literature mining and analyzing, and genotype-phenotype data through genome-wide association studies (GWAS) have enriched artificial intelligence methods and solutions, and this has allowed health providers to give personalized care by precision medicine [11].
Artificial intelligence is a branch of computer science that makes intelligent machines that behave intelligently like humans. Intelligent systems can understand complex situations, simulate human thinking and reasoning, and solve complex problems. [12]. Recent advances in the field of AI and machine learning methods have enabled them to be used in biomedical sciences and health care [13]. AI uses a set of theories, algorithms, and computing powers to perform intelligent tasks such as decision-making, reasoning, language understanding, speech recognition, and visual perception [14]. AI can increase the speed of data analysis and accuracy of decision-making in the medical area [13]. Yet, using AI algorithms in precision medicine to predict, diagnose, and treat cancer is relatively new.
1.1. Objectives. The main objective of this study is to review the applications of AI algorithms and their effectiveness in personalized medicine approaches. This systematic review tries to respond to the main subsequent questions: RQ1: What are the applications of "AI neoplasms personalized medicine"? RQ2: Which AI techniques or intelligent methods have been applied in cancer precision medicine? RQ3: In which category do each of the AI approaches fall? RQ3: How successful AI methods have been reported to improve the care of neoplasms patients?
Performing this systematic review will give researchers a broad perspective on applying various artificial intelligence techniques in personalized medicine. Also, by examining the effectiveness of different artificial intelligence techniques, researchers can select techniques that have been highly accurate in personalized medicine. This study will also introduce software and data sources used in personalized medicine for cancer. They can also have a broad view of personalized medical applications in diagnosing and treating cancer.

Methods
The following Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRISMA) for 2020 proposed by Page et al. were used in this study [15].
2.1. Eligibility Criteria. SPICE is a useful tool (like PICO) for asking focused clinical questions and qualitative reviews. The acronym SPICE stands for Setting, Perspective, Intervention, Comparison, and Evaluation and presents a way to formulate practice questions for finding evidence in existing research. SPICE may be more appropriate for formulating our research questions: Studies with the following inclusion and exclusion criteria were included in this review.

Inclusion Criteria.
The studies that met all the following criteria were entered in the review: (1) Original articles and proceedings (2) The system was designed for diagnosis, prediction, risk assessment, treatment, or screening of cancers (3) One of the AI methods was used for modeling (4) The diagnostic accuracy of the system was reported (5) The genomic, radiomic, proteomic, or phenotype data were applied in AI methods (6) Articles with English language (7) Papers that examined human-related neoplasms (8) All related studies without time limitation 2.1.2. Exclusion Criteria. The exclusion criteria were as follows: (i) The results of system test were not reported quantitatively (ii) Other than journal articles and proceedings such as review papers, letters, and book chapters (iii) The papers whose English full text of them was not available (iv) The studies whose knowledge modeling approach was not explicitly explained  December 29, 2021; we did not set a time limit for retrieving articles. Also, we searched Embase database until January 10, 2020. The search strategy used in this study included a combination of keywords and Medical Subject Headings (Mesh) terms related to "neoplasm," "precision medicine," and "artificial intelligence." Table 1 shows the complete list of keywords and terms used in the search strategy for Scopus database. A reference manager software (EndNote X8, Thomson Reuters) was utilized to collect references and to exclude duplicates. The dates of coverage for each database is given in Table 2.

Study Selection.
In this stage, assessment of records was done by more than one reviewer. The titles and abstracts of the identified articles were independently screened by two reviewers (S.R and S.S). The full text of the articles was retrieved and examined if it was supposed potentially relevant by two reviewers. Any disagreement between the reviewers was resolved by discussion by the third researcher. The following data were extracted from the selected studies and entered into a structured form in Excel. Data were extracted by each of the reviewers, and then, the forms were compared with each other. The screening procedures are displayed in Figure 1 based on the 2020 PRISMA method. Meanwhile, the main classification of reviewed articles was determined by two authors independently. The two authors (S.R and S.S) analyzed and synthesized the main characteristics of selected papers, and then, they extracted the main specification of papers. The next author (S.RNK) evaluated the extracted information and validated the main elements.

Data Collection
Process and Data Items. The first reviewer (SR) gathered the required information from the selected studies. Then, a second reviewer (SS) verified the accuracy of the information accumulated. Any dissensions were examined and resolved with a third reviewer (S.RNK). The main data elements and specifications of selected papers are displayed in Figure 2.

Study
Risk of Bias Assessment. The Joanna Briggs Institute (JBI) critical appraisal checklist for analytical crosssectional studies was used to assess the risk of bias of studies. The purpose of this appraisal is to assess the methodological quality of studies and has eight questions in the following order: (1) Were the criteria for inclusion in the sample clearly defined (2) Were the study subjects and the setting described in detail (3) Was the exposure measured in a valid and reliable way (4) Were objective, standard criteria used for measurement of the condition (5) Were confounding factors identified (6) Were strategies to deal with confounding factors stated (7) Were the outcomes measured in a valid and reliable way (8) Was appropriate statistical analysis used These questions can be answered with four options: (1) yes; (2) no; (3) unclear; and (4) not applicable.
Each "yes" answer corresponds to one score, and if 70% of the questions answered "yes" in a study, the risk of bias was considered "low." If 50% -69% of the questions were answered yes, the risk of bias was considered "moderate," and below 50% considered "high risk [16]." The checklist was completed by two authors (SR and SS), and in case of disagreement between the two authors, the disagreement was resolved through discussion with the third author (S.RNK).
2.6. Data Synthesis and Analyses. In our review, metaanalysis was not performed as the methodology and methods of reporting results in included studies were heterogeneous. The results of selected studies had been reported by descriptive statistics.

Sensitivity Analyses Conducted to Assess
Robustness of the Synthesized Results. We reviewed studies on the effectiveness of artificial intelligence techniques in personalized medicine, and the performance of which must have been quantitatively expressed. Studies with a low risk of bias were also included in the analysis. Studies should also have used cancer data to evaluate performance.

Assess Risk of Bias due to Missing Results in a Synthesis.
In this study, we had no missing results and no risk of bias due to missing results.
2.9. Processes Used to Decide which Studies Were Eligible for each Synthesis. In this systematic review, we compared and synthesized the results of studies in which the performance of artificial intelligence techniques was presented quantitatively and in the form of accuracy, precision, recall, specificity, sensitivity, AUC (area under the ROCcurve), F-score, positive predictive value, negative predictive value, Mean Average Error (MAE), and Mean Square Error (MSE). Also, due to the complexity and a large number of types of artificial intelligence techniques, we classified them into several categories, which included the following: linear model, nonlinear model, rule-based system, NLP, deep neural network, neural network, and the Bayesian model. After classifying the types of artificial intelligence techniques in these categories, we performed the syntheses. The various beforementioned indicators are defined in the following as equations: where y i = prediction, x i = true value, and n = total number of data points.
The ROC curve is constructed by plotting the true positive rate (TPR) versus the false positive rate (FPR) in diverse threshold sets. It is ideal for maximizing the TPR while minimizing the FPR. This means that the top left corner of the plot is the ideal point (FPR = 0 and TPR = 1).

Study Selection.
A total of 1788 relevant articles resulted from the search until December 29, 2021. After removing the duplicates, 1101 articles remained. Hence, in the last phase, only 63 papers that met the inclusion criteria were reviewed. In Table 3, a summary of the main results and characteristics of the papers is illustrated. Outcome measurements including results and effectiveness are summarized in Table 4. The main keywords used for selecting the papers are displayed by a word cloud scheme. Figure 3 presents the more weighted and frequent key terms used in the search. In this figure, the notion of keywords is demonstrated.

General
Characteristics of the Included Studies. The reviewed papers are presented in Figure 4 based on publication country. Forty percent of all papers were conducted in the USA, and 22% in China. The other remained countries had a relatively equal number of published articles. The frequency of selected articles based on their publication type for each year is displayed in Figure 5. The papers included in this review had been published between 2007 and 2021. As it is seen, a large number of papers had been published in 2020, 2019, and 2018. Meanwhile, the articles published in conferences are less than these, which have been presented in different scientific journals.
3.3. Source of Data and Sample Size. The selected papers had mostly publicly available data. These sources included gene expressions, gene sequencing data, phenotyping data, and somatic and mutation data. However, the molecular interactions, drug chemical data, radionics data, and pharmacogenomics data were stored in the sources. Out of 63 articles, in 19 papers, The Cancer Genome Atlas (TCGA) was used as the source of datasets. However, in 14 articles, the Cancer Cell Line Encyclopedia (CCLE) and Genomics of Drug Sensitivity in Cancer (GDSC) sources were employed as the source of datasets. In five papers, the required datasets were extracted from medical and electronic health records, and in six articles, the public websites and recruited data in papers were applied as the source of datasets. Ultimately, some other types of sources were applied in the remained papers. These sources are illustrated in detail in Table 3. Out of 63 papers, in 38 papers, the sample size was reported as patient samples, but in 16 reviewed papers, the sample size was reported as biosamples like genes, molecular samples, and cell lines. These types of samples have a large number of sizes. The reported sample size in 32 papers ranged from 30 individuals to 26,000, and also in some papers, the sample size was not mentioned.

The Distribution of Selected Papers Based on Applied
Software. Out of 63 citations, in 21 articles (33.3%), the software or technical environment was not reported, whereas, in 12 papers (19.4%), the popular and frequently used software was the R package. However, in nine of the reviewed articles, MATLAB has been used alone or in combination with other software for analysis. Also, in 10 citations (15.87%), the popular software used in the study included Weka, Python Scopus "TITITLE-ABS-KEY (("Neoplasia" OR "Neoplasia" OR "Neoplasm" OR "Tumors" OR "Tumor" OR "Cancer" OR "Cancers" OR "Malignancy" OR "Malignancies" OR "Malignant Neoplasms" OR "Malignant Neoplasm" OR "Neoplasm, Malignant" OR "Neoplasms, Malignant" OR "Benign Neoplasms" OR "Neoplasms, Benign" OR "Benign Neoplasm" OR "Neoplasm Benign") AND ("Artificial Intelligence" OR "Computational intelligence" OR "machine intelligence" OR "Computer vision systems" OR "Knowledge acquisition "OR "Knowledge representation "OR "Computer reasoning") AND ("Precision Medicine" OR "Personalized medicine" OR " Individualized medicine"))    Figure 2: The main specifications of selected papers.                 (i) The MPS system is simple and cost-effective to apply and yet can reveal previously unrecognized heterogeneity among patient subpopulations in a platform-independent manner.
(i) The integrative nomogram incorporated CECT radiomics, transcriptomics, and clinical features improved the PFI prediction in BLCA patients and is a feasible and practical reference for oncological precision medicine.
(i) The radiomics signature model achieved a better classification performance than radiologists, which demonstrated the impressive prediction ability of radiomics signature.
[79] Accuracy of 81% and AUC of 0.896 for the ROC curves.
(i) The model exhibited good interstage prognosis prediction performance. The genetic features could be used as biomarkers for effective LUAD prognosis prediction [51] Accuracy: (i) The proposed methods were able to distinguish the metastatic sclerotic lesions with a complete response.
[70] Accuracy = 0:9143: (i) This synergy between liquid biopsy biotechnology and XAI will surely lead to personalized interpretable medicine, ensuring adequate and better diagnostic tools and treatments.
[60] AUC = 0:97 − 0:98: (i) A PLATYPUS model trained on the drug trial data can predict drug response for this patient without retraining.
(i) This was concluded that this deep learning model provides an accurate and reproducible method for the prediction of BAP1 expression in uveal melanoma. [25] Sensitivity: upper than 84% in the training set but below 77% in the testing set.
(i) This study demonstrated that MRI-based radiomics features hold potential in the pretreatment prediction of response to NACT in LACC, which could be used to identify rightful patients for receiving NACT avoiding unnecessary treatment.
(i) The TCPR model may benefit decision-making regarding total laryngectomy or larynx-preserving treatment. This TCPR model incorporating radiomics signature and T category reported by radiologists has good potential to be applied for individual accurate preoperative T category. (i) The proposed framework had demonstrated its capability and potential for mapping the gene and tumor status, it was effective for detecting association between gen information and the tumor growth regions.
(i) CDRscan is expected to allow the selection of the most effective anticancer drugs for the genomic profile of the individual patient.
(i) The result shows that GloNetDRP achieves comparable performance on the two-omics data for eight drugs collected from CCLE and GDSC. GloNetDRP globally calculated the responses of untested cell lines for the query drug by considering not only the neighbors but also other drugs and cell lines.

BioMed Research International
The experiments show that the OFSSVM is an appealing compromise between interpretability and classification accuracy, and is superior to other traditional methods in the sense of comprehensive evaluation.
[72] FPR for DNT and DMT p values at α = :05 for Sc1: 0.04 and 0.208 (i) PANOPLY can be a tool to help clinicians in their decision-making process.
(i) The proposed approach surpasses the prognostic accuracy of human experts for classifying brain tumors. [44] ROC curve of the Gemelli polyclinic's data set = 0:759. ROC curve of the Maastricht clinic = 0:881. ROC for the testing set was depicted 0.603 and 0.588 for each data set.
(i) Experimental results indicate that the system can generate a highly performant center-specific predictive model. [73] Accuracy across all pathways was 0.96 for a single dataset and 0.72 with multiple datasets (i) GRAPE pathway scores provide researchers with a unique perspective of patient transcription profiles that may lead to improvements in the prediction performances of a wide range of personalized medicine applications. [54] For CTRP panel, the median was calculated for GBGFA, ENET 0.05, and 0.04. For CCLE panel, the median was calculated for GBGFA and ENET 0.06, 0.02.
(i) Current results show that the GBGFA model enables leveraging information from combinations of genomic data which improves the predictive performance and feature selection as compared to Elastic Net and BGFA.
(i) The results suggested the effective therapies for the majority of cancer cells investigated in the dataset.
(i) This QA system can be effective for helping physicians in relevant knowledge. So, precision oncology can provide fewer toxic treatments in neoplasms.
[78] More than 90% accuracy (i) The analysis demonstrated that voting of the output categorical values for a given patient across distinct prognostic/classification methods could lead to a more robust, accurate, reproducible, and cost-efficient prognostic/ classification strategy for precision medicine.
(i) The proposed algorithm improves the cost efficiency and accuracy of the screening process compared to current clinical practice guidelines.
(i) The proposed approach has the potential to enable the derivation of new hypotheses, improve drug selection, and lead to an improvement in patient genomics-tailored therapeutics for cancer.
[71] Range of AUC 0.58-0.64 (i) Such studies are expected to contribute to precision medicine and better guide treatment for these deadly diseases [74] (≥0.80 % accuracy for 10 drugs, ≥75 % accuracy for 19 drugs (i) This model could be applied to predict drug response for some certain drugs and potentially play a complementary role in personalized medicine. [29] Sensitivity from 90% to 95%, specificity 67% to 93%).
(i) This study opens the way to further development for identification of new biomarker combinations in other applications such as prediction of treatment response.
(i) The ability of Afirma BRAF to accurately detect V600E status may assist physicians in making these treatment decisions and potentially improve patient care.
(i) In this study, the CANScript platform was versatile in its ability and capacity to predict the outcomes of both cytotoxic chemotherapy regimens and targeted therapeutics.
(i) The findings suggested that decision trees and support vector machines are engaged approaches for clinical decision support in the patient selection for targeted therapy in advanced NSCLC. [67] Average accuracy for leukemia: 92.90%; breast 84.67%; colon cancer 86.53%; (i) The proposed two-step Bayes classification framework was equal to and, in some cases, outperformed other classification methods in terms of prediction accuracy, the minimum number of classification markers, and computational time. (i) This model had been shown that the prediction of drug response and mode of action by transcriptional profiling is significantly and effectively enhanced when paired with known a priori gene and protein networks.
The average training accuracy of 0.6995 and average testing accuracy of 0.6042.
(i) This investigation implicated XPD 751, XPD 312, and pack-years of smoking as significant predictors of bladder cancer susceptibility. [46] This algorithm performed better than simple metrics for variation in individual and multiple genes (R 2 = 0:10; p < 0:05).
(i) This approach performed better than simple metrics for variation in individual and multiple genes [62] Average accuracy for SVM, NBC, and BCN was calculated respectively 85.7, 94.2, and 94.3.
(i) As this contribution, the experiments with lung cancer data prove that RPPA data can be used to profile patients for drug sensitivity prediction.
(i) The author believed that this method can be a useful tool for translating the gene expression signatures into clinical practice for personalized medicine.
(i) The C-T CERP algorithm appears to have a good potential and effective role for biomedical decision making in the assignment of patients to treatment therapies. (i) This study demonstrated that the CECT-QC algorithm is useful for radiomic-based precision diagnosis [81] The trained DL model classified patients into high-risk and low-risk groups in training cohort (p value < 0.001, concordance index (Cindex): 0.82, hazard ratio (HR): 9.79) and external validation cohort (p value < 0.001, C-index: 0.78, HR: 11.76).

BioMed Research International
(i) DL model can provide CT-based prognostic risk scores related to the OS of GC patients, and the findings demonstrated higher prognostic value than clinical and radiomics models.
(i) The results show quite a high prediction accuracy, which proves the discriminative ability of the proposed model. (i) The proposed R-SNN maintains crucial features by using the residual connectivity from the encoder to the decoder, and it also uses only a few layers, which reduces the computational cost of the model. [84] The accuracies of training, validation, and test dataset were 93.5, 93.7 and 98.1%, respectively; AUROC value of 0.98 was observed for both the classes.
(i) The proposed omics integration strategy provides an effective way of extracting critical information from diverse omics data types enabling estimation of prognostic indicators.
(i) The decision tree exceptionally demonstrated high classification precision with cancer data, producing consistently relevant forecasts for the sample tests with an accuracy close to the best ones achieved from supervised ML algorithms. [86] On the GDSC dataset, the AUCROC of RefDNN were 0.891; the AUCROC of RefDNN were 0.071 on the CCLE dataset.
(i) As the proposed model can guarantee good prediction of drug responses to untrained drugs for given gene expression patterns, it may be of potential benefit in drug repositioning and personalized medicine [87] The median AUC value per target pathway ranges from 0.98 for hormone-related drugs to 0.73 for compounds targeting metabolism pathways.
(i) Appropriate feature selection strategies facilitate the development of interpretable models that are indicative for therapy design.
(i) The results demonstrate proposed framework improves the prediction performance in all three drug response prediction applications with all three prediction algorithms. [88] Average accuracy of ECF-W and ECF-S is 74.25% and 77.25%, respectively.
(i) These two methods recommend the most suitable compounds and anticancer drugs for patients with NSCLC.
[ 3.5. The Characteristics of Selected Articles Based on Input, Cancer Types. First of all, it can be said that the inputs of applied AI methods and algorithms were categorized into genomic data (gene expression, somatic mutation, phenotype data, sequencing data, and proteomic data), functional data, and radiomics data (radiogenomic biomarkers and histology of images). Based on literatures, the genomic data include profile of DNA, proteomics, transcriptomics measure transcripts, metabolomics, and radiomics. These concepts have created multiomics new profiles. Radiomics is an approach that extracts a large number of features and critical characteristics from radiographic images. Hence, detecting correlations with genome patterns is mining if radiomic data is known as radiogenomics (59,60). Out of 63 papers, in 34 articles (53.9%), genomic data like gene expression, somatic mutation data, phenotype data, proteomics with drug-response data was used as input in AI methods; in 12 citations (19.4%), radiomic data (radiographic with biomarkers) was applied by researchers for managing neoplasm' treatments. However, in 16 papers (25.3%) drug response, functional data was applied in per-sonalization of treatment. drug response which is functional data. Concerning the type of cancer, in seven papers out of 63, the selected cancer for treatment (predicting, diagnosing, and treating) was breast cancer, and the data of lung cancer was used in seven reviewed papers. Also, in four citations, the data of ovarian cancer was used by the researchers. In some papers, the data of cancers such as bladder cancer, thyroid cancer, colorectal cancer, brain tumor, laryngeal carcinoma, leukemia, and neck cancer were utilized too. The type of selected cancers and type of inputs (source of omics data) are displayed in detail in Figure 6. However, in Figure 7, the distribution of papers by the type of care and input is depicted.
3.6. Distribution of Selected Papers by Effectiveness. The effectiveness of selected AI methods in various reviewed papers is shown in Table 5. The results showed that AI algorithms have the potential and capacity to predict, diagnose, and treat cancer (drug-chemotherapy evacuation, etc.). These methods can classify or stage patients and provide better therapy measures. The performance of applied methods was validated and evaluated by different beforementioned criteria, including accuracy, sensitivity, specificity, AUC of ROC, Mean Average Error (MAE), and F-measure. Maximum values of the assessment indicators such as accuracy, sensitivity, specificity, precision, recall, AUC, F-score were calculated to be 0.99, 1.00, 0.96, 0.98, 0.99, and 0.9929, 0.98, respectively. These reported criteria showed that the performance of the methods was at a significant level. Hence, many of the algorithms proposed in reviewed studies have effectively performed early detection of cancers, predicting response to treatment, and screening through personalized medicine.

The Distribution of Citations Based on the Type of
Presented Care by AI Methods. Based on extracted results, in 41% of studies, the main purpose of using genomic data in artificial intelligence was to predict the response to drugs in the treatment of cancer patients and in 12.7% of the papers, correct diagnosing of neoplasms by AI approaches was the critical care. The offered type of care in studies is displayed in Figure 8.  The main objective of this review was to determine the application of AI techniques in precision medicine for cancer screening, diagnosis, and treatment. In Table 4, an overview of the distribution of applied AI algorithms, their categorizations in the selected papers, and the frequent methods used in the reviewed papers are presented. Out of 63 citations, the leading AI approaches in 17 papers (26.9%) include linear and nonlinear models (classification and regression trees, support vector machine, Neural Networks, and etc.). In 15 articles (23.8%), unique methods based on artificial intelligence were used, considered linear models (Random Forest or Decision Trees). Totally, in 21 citations, rule-based systems and deep learning models were used too. Some other intelligent techniques such as metaensemble feature selection, kernel learning, natural language processing, and the least absolute shrinkage and selection operator Cox regression were employed by researchers to determine cancer characteristics and input methods.
3.9. Risk of Bias within Studies. Sixty included studies in this review were considered low risk of bias. Only two citations were evaluated with moderate risk of bias [18,19] and one with high risk of bias [20]. The questions "Were confounding factors identified? and "Were strategies to deal with confounding factors" are not applicable in our included studies, because our studies were not experimental researches.

Discussion
According to the results, the leading artificial intelligence methods and applications are widespread and lead to knowledge-based production or model development applied widely in healthcare fields. The main objective of this review was to analyze and identify the studies conducted on the application of AI methods in precision or personalized medicine for cancer prediction, diagnosis, and treatment. To achieve this objective, we selected 63 papers based on inclusion and exclusion criteria. The basic aim of reviewed studies was to provide or propose AI-based approaches that   More than one type of cancer  Figure 6: The distribution of papers based on inputs and type of cancer.

BioMed Research International
can predict the outcomes of treatments such as drug therapy or chemotherapy and patient screening/diagnosis. However, it should be noted that the type of care in a large number of selected citations was predicting the type of treatment based on the stage of cancer and identifying the mutations and sequencing in genomics. Support vector machine is one of the most preferred methods of machine learning that has high accuracy. This algorithm is capable of handling a large volume of data [21], and it is also a method of choice when dealing with large and complex data as it can provide statistical analysis and summarization [22]. Another algorithm employed in the reviewed papers was the random forest algorithm, which can be used widely in more applications, specifically with large datasets.
Hence, based on effectiveness reports, in large numbers of the selected papers, the SVM-based and RF methods effectively predicted and diagnosed cancer with genomic data. Another algorithm that was used in radiomicoriented papers was convolution neural network (CNN), which is a deep learning technique that can take in an input image and is designed to improve the accuracy of automatic labeling and classification [23,24]. In some reviewed papers, the researchers used CNN to propose and predict cancer patients based on radiogenomics and histology images. Based on our results, the qualitative criteria such as ROC, accuracy, sensitivity, specificity, precision, and F-score were reported separately for reviewed papers. For example, in a study in 2019 that examined the response of drug and chemotherapy in patients with cervical cancer, the sensitivity of SVM models was higher than 84% and AUC was 0.99 for testing the set [25]. In 2018, Brietenstein et al. proposed a rule-based algorithm for diagnosing and treating patients with breast cancer, which had the precision of 0.98, recall of 0.99, and F-score of 0.98 [26]. Meanwhile, Chen et al. proposed a radiomic-based model for selecting proper surgical approaches in papillary thyroid carcinoma; this model had an AUC of 0.837 and F-score of 0.766 [27]. In another work, an SVM-based classification model was proposed for cancer prediction, which had an accuracy of 70% and AUC of 0.9929 [28]. Similarly, Kempowsky-Hamon et al. developed a fuzzy logic algorithm for predicting breast cancer prognosis; the performance of this model was reported by sensitivity and specificity (0.95 and 0.93, respectively) [29]. Notably,   [30]. In [31], the consequences demonstrate offered ensemble transfer learning methods sweeten the prediction performance in all three drug response prediction applications with all three prediction algorithms. According to the drug response data, AUC values (0.98, 0.99) are computed and abode as the drug response measurements to be predicted through regression analysis. Nevertheless, due to the heterogeneity of reports and our results, we could not analyze them one by one. However, the selected papers showed that their methods were influential in the precision medicine field.
Our results showed that the main type of cancer was breast and lung cancer in most of the selected studies. Breast cancer is cancer that forms in the cells of the breasts. After skin cancer, breast cancer is the most common cancer in women. Based on results, breast cancer survival rates have increased, but the number of death associated with this cancer is declining [32]. However, breast cancer is the most invasive cancer in women and the second cause of death after lung cancer [6,7]. For this reason, a new personalized approach was introduced with the name of precision medicine used to diagnose, treat, and prevent the number of cancers such as breast cancer by taking into account the genes (genetic makeup) or other markers in the cancer cells [33]. In other words, if the cancer treatment and diagnosis in individuals are based on the formation of genome profiles, personalized medicine is a more effective method that can be used in the treatment process [34]. The blood or tumor tissue in this method is collected for genetic analysis to determine its genetic makeup, which may later help predict or diagnose cancer or guide the treatment decisions. Other tests determine the genetic changes or variants (called mutations) within the cancer cells. This information can help determine which treatments are most likely to be beneficial or if any treatment is needed at all. For example, cells from a breast tumor may be tested to determine whether they produce too much of a protein called HER2 [35]. Lung cancers differ according to the type of cell in which they arise. In these cancers, specific molecular targets have been identified, and which gene alterations produce mutations [36]. Therefore, if the genetic abnormality is identified, it can be targeted by a drug [37]. According to our results, genetic is very effective in the onset of breast and lung cancer, which were the most common types of cancer in the selected papers. The source of our reviewed papers was publicly available. These sources include gene expressions, gene sequencing data, phenotyping data, and somatic and mutation data. However, the molecular interactions and drug chemical data, radiomics data, and pharmacogenomics data were stored in the sources. In most of the selected papers, TCGA, CCLE, and GDSC sources were used. These three sources have the main datasets on patients or biosamples, and their data are publicly available with various types of genomics and radiomics data. The results showed that most of the articles had been published in the United States, and the number of articles published in the field of precision medicine has increased significantly in recent years. Various factors may have led to an increase in publications of such articles in recent years, especially in the United States. The term "precision medicine" was first highlighted in a publication by the US National Research Council, which sought to create a new taxonomy for classifying diseases through a knowledge network [38,39]. US National Human Genome Research Institute (NHGRI) has developed a 20-year plan for translating insights from genomics to medicine. This has led to an understanding of human biology and disease prevention, diagnosis, and treatment [39,40]. President Obama, on January 20, 2015, announced that "Tonight, I am launching a new Precision Medicine Initiative to bring us closer to curing diseases like cancer and diabetes-and to give all of us access to the personalized information we need to keep ourselves and our families healthier." This could have led to progress in precision medicine [41]. President Obama has allocated $215 million for the initial launch of this initiative. He also donated $130 million to the Cohort study, which involved at least one million volunteers. This can lead to the collection of genotypic, phenotypic, and lifestyle data and thus accelerate the development in this scientific field [42].
All of the mentioned items can be the reason for the advancement of precision medicine in the United States, especially in recent years. It seems that other countries should take a step in this direction and keep pace with the United States in developing this area by allocating sufficient budget and time.
This study had several strengths. One of the strengths of this study was searching in critical databases including Med-line (through PubMed), Scopus, Cochrane Central Register of Controlled Trials, Embase, IEEE Xplore Digital Library, and ISI Web of Science, which enabled us to cover all the articles published in this field as much as possible. Another strength of this study was the inclusion of papers presented at the conferences. We also did not impose any time limit on the search.
In this study, we faced some limitations, one of which was the noninclusion of studies presented at conferences, that we did not have access to their full texts. We also used only English papers, so there is the probability of missing several related studies and papers with effective results. However, we could not update Embase search due to lack of access to Embase database from our country.

Conclusion
This review was conducted to examine the applications of artificial intelligence in personalized medicine. To achieve this goal, we investigated five important databases to retrieve published scientific papers without time limitation. Hence, applying appropriate AI-based solutions could improve the treatment and management of cancers and the application of intelligent approaches is recommended in many areas such as in personalized medicine. However, further studies are needed to investigate the real effects of these algorithms and their effectiveness.

Data Availability
All data generated or analyzed during this study are included in this published article.