Electronic health record (EHR) data have been used to store patient-specific information for decades, including structured data, such as diagnosis, medication, laboratory test results, and unstructured data obtained from clinical reports. Observational patient data are used in vast computational analyses, including the generation of individual patient profiles and detection of patient similarity based on clinical and genomics data [1]. With advancements in the field of genetics, it is possible to analyze large amounts of genomic data using different ML and other predictive methods that can widen the knowledge about diseases with a genetic background, such as cancer, as well as rare and unclear diseases. However, such algorithms need large patient cohorts to reach a clinical prediction scale and useful diagnostic decision support [2]. For this purpose, a harmonized research data repository is necessary to enable a joint analysis across institutions based on observational data [3].

The Observational Health Data Science and Informatics (OHDSI) initiative is a promising international effort to optimize secondary use of observational data by harmonizing and standardizing clinical data and to create scalable analytical tools [4,5]. The basis for this is the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM), which ensures homogeneous storage of observational healthcare data across different databases with interoperable formats and standard terminologies. The terminologies for diagnoses/conditions, observations, and drugs within the OMOP CDM are based on, for example, International Classification of Diseases (ICD) codes [6], Systematized Nomenclature of Medicine – Clinical Terms (SNOMED-CT) [7], and normalized naming system for generic and branded drugs (RxNorm) [8]. To apply these concepts, one usually needs to retrieve the already mapped tables from the Automated Terminology Harmonization, Extraction and Normalization for Analysis (ATHENA) [9] standardized vocabulary tool from OHDSI. Afterwards, the harmonized data stored in the OMOP CDM format can be used in systematic studies, population-level estimations, drug and biomarker evaluations, as well as further patient-level prediction [10].

To develop tailor-made therapies for cancer patients, researchers must have access to genetic variants and their associated pathways together with the clinical information. Studies such as Unberath et al. [11] created a vocabulary set using the HUGO Gene Nomenclature Committee (HGNC) [12] for their specific use case, but a comprehensive standard vocabulary that can assist in sequencing data on the OMOP CDM and predictive modules in general would be essential.

The first attempt for this purpose was a Genomic Common Data Model (G-CDM) proposed in 2019 [13] to store next-generation sequencing (NGS) data. The G-CDM introduced four extension tables to the OMOP CDM which, to use any of the standardized OHDSI tools for the purpose of data analysis or prediction on data stored using this format, would require adoption of the tools, because they strictly follow the OMOP CDM structure. One important tool is named ATLAS – A unified interface for the OHDSI tools [14], which is an open-source web-based interface to configure analysis modules such as Patient Level Prediction (PLP) in the OMOP CDM [15,16]. In particular, PLP can be used to define Artificial Intelligence (AI)-based analyses on previously defined patient cohorts by using an easy-to-use graphical user interface. However, to tackle the challenge of enabling oncology data in the OMOP CDM without any structural changes, the OHDSI community has published the first version of a set of new vocabularies for presenting genomic data in the OMOP CDM in 2020 [17]. These new vocabularies are built based on different gene databases, e.g., ClinVar [18], Clinical Interpretation of Variants in Cancer (CIVic) [19], and Precision Oncology Knowledge Base (OncoKB) [20]. With the use of new vocabularies, the data can be represented in the OMOP CDM without the need for any further extensions. Additionally, an OMOP oncology module was introduced in 2021 [3], which extends the OMOP CDM and its terminologies to support the harmonized documentation of cancer conditions, treatment, and disease abstraction. This module uses the concepts from seven existing standards, namely, International Classification of Diseases for Oncology (ICD-O-3) [21], Hematology Oncology (HemOnc) [22,23], North American Association of Central Cancer Registries (NAACCR) [24], College of American Pathologists CAP [25], Nebraska Lexicon [26], National Cancer Institute (NCI) [27], and Anatomical Therapeutic Chemical (ATC) classification [28]. Moreover, the Radiology Common Data Model (R-CDM) for standardization of Digital Imaging Communications in Medicine (DICOM) was published in 2021 [29]. R-CDM uses the RadLex glossary, which contains 75,000 radiology terms to harmonize DICOM imaging data into two extended tables, radiology occurrence and radiology image, on the OMOP CDM.

Clinical integration of the OMOP CDM can pave the way back to patients through facilitating access to relevant data, enabling multicentric, multidatabase studies to enhance statistical power and transfer results across populations [30,31,32]. With the recent advancements in the field of medical informatics, many predictive algorithms are known and used in the field of oncology, which benefits largely from the use of such models in uncovering unknown information about the cause and course of certain types of cancer. For the purpose of this review, we looked for studies that have taken advantage of such predictive models to perform cancer-related analyses on an OMOP CDM and evaluate to what extent the genomic vocabulary extension of the OMOP can serve current needs of ML-based predictions.

The literature screening resulted in 248 papers from 13 search engines, of which only five matched the scope of our review and are finally included (Figure 1). In particular, after duplicates’ removal, 212 articles were title- and abstract-screened. In this step, articles that did not indicate an AI-based prediction analysis and OMOP in their title or abstract were excluded.

Afterwards, the full text screening step analyzed in total 29 articles, out of which 15 were either focused on AI and cancer but without using the OMOP or focused solely on cancer or AI. The remaining nine articles either contained cancer studies on OMOP-based data not using predictive AI models [11,33,34,35,36] or performed predictive analysis on OMOP-based data of a non-cancerous disease [37,38,39,40,41,42,43,44,45]. An example for the first group are preliminary studies that are focused on harmonizing data in the OMOP using extract, load, and transform (ETL) processes. The articles that perform predictive analysis on other than cancerous data partially use different machine learning and deep learning methods. One of these studies is Hardin et al. [46] that uses the OHDSI PLP module for the development of predictive models. Since these excluded studies also contain a valuable source of information for the current review, detailed information of the most important excluded articles and the finally included five articles can be obtained in the attached Supplementary Table S1 (color-coded in grey). In the following, we highlight the studies that ultimately contain aspects of AI and the OMOP in the cancer domain.

Among the included papers, Felmeister et al. [1] focus on the pediatric rare brain tumor and follow an exploratory approach to extract pertinent information from a large simulated observational dataset based on the OMOP and discover data points that contribute to the data-driven phenotype of a diagnosed subject. An example of such a data point is population-based survival estimates. The authors apply a supervised prediction approach and take advantage of the Logistic Regression (LR), Linear Discriminant Analysis (LDA), K-Neighbors classifier (KNN), Decision Tree classifier (CART), Gaussian Naïve Bayes (NB), and Support Vector Machine (SVM) algorithms. The models are applied on a simulated cohort of 1000, in which KNN performs best with the highest percentage of correctly identified cases. SVM and LR are the second and third best-performing ML algorithms. The analysis shows that usage of OMOP CDM observational data in exploration analysis can lead to valuable discoveries.

Meystre et al. [47] train a Natural Language Processing (NLP) method using manually annotated physician letters for subsequent automatic detection of patient eligibility for breast cancer clinical trials. The authors encode the clinical trial eligibility criteria to the corresponding clinical information system. The clinical notes were stored in the notes table in the OMOP CDM. They use NLP and an SVM classifier method to extract the patient-derived EHR information from the existing free text notes written by physicians. The cohort was designed by using the ATLAS platform of OHDSI. The study shows that NLP is able to extract the eligibility criteria for clinical trials from EHR notes from a cohort of 229 patients, with an average recall and precision of 84.6% and 64.4%. In comparison, SVM models perform better with an average recall of 90.9% and precision of 89.7%. Using the extracted eligibility information, the patients were classified to determine eligibility using an SVM binary classifier with high accuracy.

Unlike Felmeister et al., the third study by Seneviratne et al. [48] uses tree-based classification models, such as Lasso Penalty (LASSO), Random Forest (RF), Gradient Boosted Machine (GBM), and Extreme Gradient Boosting (XGB), on a cohort with prostate cancer. The algorithms classify metastatic cancer from non-metastatic cases based on the stage of cancer, which is usually documented in text form in medical notes, which means it is only feasible to extract cancer stage information on population level, when an AI-based approach is used. The study demonstrates identification of patients with metastatic prostate cancer in a cohort of 5861 patients using an RF classifier with a precision and recall of 90% and 40%, respectively. The RF model outperforms other models, including normal ICD code search, which leads to a recall and precision of 54% and 33%.

Moreover, the Information Technology for the Future of Cancer (ITFoC) [49] introduces a framework for the validation of AI algorithms with omics and clinical data for prediction of the treatment response in triple-negative breast cancer (TNBC) [50]. In this framework, the AI models will be developed and validated on real-world data. The clinical and -omics data will be harmonized via the OMOP CDM and terminologies, such as ICD-10, Logical Observation Identifiers Names and Codes (LOINC), and SNOMED-CT.

Furthermore, Lee et al. [51] perform a retrospective study of data obtained from seven hospitals in Korea that adopted the OMOP CDM as main research data repository. The study aims to find the association of angiotensin-converting enzyme inhibitor (ACEi) and angiotensin receptor blocker (ARB) with lung cancer development. Similar to Meystre et al., for cohort definition and defining the baseline characteristics of the study, the OHDSI tool ATLAS was used.

As shown in Table 1, all of the abovementioned articles use the OMOP CDM as a data standardization model, transform their datasets to this format, and design their AI-based analysis on it. The vocabularies that are used in these papers to transform data into the OMOP CDM structure include the International Classification of Diseases Clinical Modification, 9th Revision (ICD-9-CM), International Classification of Diseases Clinical Modification, 10th Revision (ICD-10-CM), SNOMED-CT, and LOINC codes. Moreover, a wide range of models are used as predictive models in the aforementioned papers, starting from classical machine learning methods, e.g., RF, GBM, all the way to other regression and classification methods, including linear regression, lasso regression, SVM, and k-Nearest Neighbors (KNN). Since the OMOP CDM harmonizes different data structures, the same predictive tool or trained model can be applied in different medical studies.

In summary, all of the abovementioned finally included articles use the OMOP as a common data model, out of which two (Meystre et al. and Lee et al.) use ATLAS for the purpose of cohort definition. Seneviratne et al. use only tree-based methods, whereas the other studies commonly utilize a combination of different types of methods (e.g., tree-based, boosting, SVMs). The use of different methods can also be obtained by using the PLP from OHDSI. Only Felmeister et al. use Centers for Medicare and Medicaid (CMS) Medicare Claims Synthetic Public Use Files (SynPUF) simulated data [52], which is a freely available dataset converted to the OMOP CDM used for benchmarking studies and technology implementations. A single study (Meystre et al.) uses unstructured, i.e., free text data for the initial analysis, and Tsopra et al. uses -omics data in addition to structured clinical data.