This research was conducted as part of the Psychiatric Genomics Consortium’s (PGC) ongoing effort to identify and collect DNA samples from patients with OCD, using 2 EHR-linked biobanks within the PsycheMERGE Network: Vanderbilt University Medical Center’s biobank (BioVU) at VUMC in Nashville, Tennessee, and the MGB Biobank (MGBB) [23] at MGB in Boston, Massachusetts. The study included 213,275 individuals receiving care at VUMC from 1989 to 2021, and 97,275 patients at MGB from 1976 to 2021. Additional data details are provided in Multimedia Appendix 1.

Ethical approvals were obtained from the institutional review boards at both VUMC and MGB, including an informed consent waiver for the use of retrospective medical record data without patient interaction. All participants provided written informed consent for their inclusion in the respective biobanks for broad-based research use.

Two reviewers with clinical training in psychiatric assessment (EM: a postpsychiatry clerkship medical student; PM: a psychiatry resident) were selected to conduct chart reviews at VUMC. Both were trained using chart review instructions (Appendix S4 Multimedia Appendix 1), which included Diagnostic and Statistical Manual of Mental Disorders (Fifth Edition; DSM-5) diagnostic criteria for OCD, evidence grading guidelines, and defined inclusion and exclusion rules. OCD cases for chart review were identified using the ICD-or-NLP definition. Both reviewers independently examined charts from the same set of 25 subjects to assess interrater reliability (IRR). Reviewers, blinded to each other’s assessments, determined OCD presence and graded supporting evidence. A psychiatrist (JB) and a psychiatric nurse practitioner (HH), both with clinical expertise in OCD, adjudicated discrepancies and provided feedback. Once adequate agreement (IRR>90%) was achieved, each reviewer independently assessed an additional 25 subjects. If agreement was inadequate, another 25 subjects were reviewed, followed by the same confirmation process until the required agreement threshold was met. Once confirmed, the remaining subjects were reviewed independently.

For external validation at MGB, 2 board-certified psychiatrists (JR and RK) with extensive experience in psychiatric chart review conducted 2 review rounds for algorithm refinement and evaluation. Disagreements were adjudicated by a third experienced psychiatrist (JS). The first 2 rounds together included 68 algorithm-identified cases and 8 noncases. In the final round, 3 clinical psychologists (EG, RF, and SW) reviewed 90 randomly sampled algorithm-identified cases. All 3 reviewers first independently reviewed the same 10 charts for calibration, followed by a meeting to align on diagnostic criteria and evidence grading. Each reviewer then independently assessed 33‐35 charts, with each chart reviewed by 2 reviewers and discrepancies resolved by the third. Reviewers at both sites followed standardized chart review instructions (Appendix S4 and Figure S2 in Multimedia Appendix 1).

For genetic analysis, we used BioVU, VUMC’s biobank that links deidentified DNA samples with EHRs. BioVU, established in 2007, obtains patient consent through forms provided in outpatient clinic settings. We applied quality control measures (described in Appendix S5.1 Multimedia Appendix 1) to the algorithm-defined OCD cases and controls in BioVU to generate the final dataset used for genetic analysis. OCD cases were matched to controls at a 1:3 case-control ratio based on median age in the medical record, sex, and the first 2 genetic principal components.

The OCD PGS was derived from the OCD GWAS by Strom et al [14] (excluding 23andMe and BioVU samples). The PGS for height and major depressive disorder (MDD) were derived using GWAS summary statistics for height [27] and MDD [28], respectively. We used Polygenic Risk Score via Continuous Shrinkage (PRS-CS) [29] to estimate posterior single-nucleotide polymorphism (SNP) effect sizes, using CEU (refers to Utah residents with Northern and Western European ancestry) samples from the 1000 Genomes Project Phase 3 as the reference panel. PGSs were calculated as the sum of the risk allele dosages weighted by the posterior SNP effect sizes. Logistic regression assessed the association of the OCD, MDD, and height PGSs with algorithm-defined OCD status, adjusting for current age, genotype batch, and principal component analysis components 1‐10. Nagelkerke R² was calculated for the full model and for the corresponding model excluding the PGS term, and the difference was reported to quantify the incremental contribution of each PGS to OCD case status. Additionally, we performed a sensitivity analysis that included MDD status as a covariate to assess whether associations between the OCD, MDD, and height PGSs and OCD case status were independent of comorbid MDD. MDD status was extracted from the EHR using a previously established algorithm [30,31], as detailed in Multimedia Appendix 1.

The MGBB, established in 2010, collects biological samples, health information, and genetic data to support biomedical research [23]. Participants are enrolled using a broad-based consent process by research coordinators at clinical and public hospital locations or electronically through the MGB patient portal. Genotyping was performed for 53,853 participants in 13 batches using 4 Illumina array platforms: Multi-Ethnic Genotyping Array (MEGA, batch 1), Expanded Multi-Ethnic Genotyping Array (MEGAEX, batch 2), Multi-Ethnic Global (MEG, batches 3‐9), and Global Screening Array (GSA, batches 10‐13). Batches 8 and 9 were excluded due to few OCD cases (see Table S1 in Multimedia Appendix 1 for case and control counts). Quality control procedures are detailed in Appendix S5.2 and Figure S4 in Multimedia Appendix 1.

Consistent with methods used at VUMC, SNP effect sizes from GWAS of OCD, MDD, and height were used to construct PGSs. Posterior SNP effect sizes were calculated using PRS-CS [29], with the CEU reference panel. PGSs were calculated as weighted sums of risk allele dosages based on the posterior SNP effect sizes. Logistic regression analyses were performed separately in the MEG and GSA datasets to evaluate associations between each PGS and algorithm-defined OCD case status, adjusted for age and genetic principal components. The incremental contribution of each PGS to OCD status prediction was assessed by calculating the difference in Nagelkerke R² between the full model and the corresponding model excluding the PGS term. Sensitivity analyses were conducted by including MDD status as a covariate to assess whether the associations of the OCD, MDD, and height PGSs with OCD case status were independent of comorbid MDD.

At VUMC, reviewers achieved 91% raw agreement and Cohen κ=0.72 in the first 2 rounds (50 charts total), indicating substantial IRR [32]. In the third and final round, each reviewer independently assessed 22‐25 additional charts, totaling 97 reviews. At MGB, 2 initial rounds of chart review identified 34 OCD cases and 42 noncases, with reviewers agreeing on 71 out of 76 charts (93% agreement; κ=0.91). In the final round, 3 reviewers (EG, RF, and SW) assessed 90 charts (52 cases and 38 noncases), achieving 87% agreement and κ=0.71. At MGB, only the final round was used for algorithm validation (Table 1).

Table 1 summarizes chart review validation results. The ICD-and-NLP definition (intersection) achieved the highest PPV of 0.84 at VUMC and 0.91 at MGB, but identified fewer cases (515 and 335, respectively). The ICD-or-NLP definition (union) yielded more cases (1057 at VUMC; 1379 at MGB) but lower PPVs (0.72 and 0.58), with 38 of 90 MGB subjects misclassified as OCD cases. These findings highlight the tradeoff between PPV and the number of identified OCD cases when comparing the disjunction (ICD-or-NLP) and conjunction (ICD-and-NLP) of the 2 arms in the algorithm. The NLP-only definition yielded the lowest PPVs at both sites. Replication at MGB resulted in lower PPVs for 3 of the 4 definitions (ICD-defined, NLP-only, and ICD-or-NLP), particularly the latter 2, underscoring the challenges associated with using clinical notes across different institutions.

Among 213,275 BioVU participants, our algorithm identified 640 genotyped OCD cases by ICD codes (ICD-defined), 417 by NLP (NLP-only), 1057 by either arm (ICD-or-NLP), and 515 by both (ICD-and-NLP). For MGBB, 3385 of 97,275 participants were excluded based on predefined criteria (Figure 1A). Of 93,890 remaining, 53,853 underwent genotyping, yielding 725 ICD-defined, 654 NLP-only, 1379 ICD-or-NLP, and 335 ICD-and-NLP cases (Figure S3 in Multimedia Appendix 1). Pie charts summarizing the OCD case counts are presented in Figure S3 in Multimedia Appendix 1.

Table 2 and Table S2 in Multimedia Appendix 1 summarize demographic characteristics. ICD-defined cohorts included older participants (mean age 46.2, SD 17.53 at VUMC; mean 55.4, SD 17.58 at MGB) and a higher proportion with public insurance (55.2% and 65.4%). In contrast, NLP-only cohorts were younger (44.0 at VUMC; 45.2 at MGB) with fewer on public insurance (45.6% and 50.8%). Females predominated across all cohorts, especially NLP-only (65.7% at VUMC; 67.3% at MGB). Most individuals self-reported as White (91%‐92%; at VUMC; 83%‐91%; at MGB), highest in ICD-and-NLP cohorts. In terms of hospital utilization, NLP-only cohorts consistently had the fewest visits, diagnoses (ICD), procedures (current procedural terminology [CPT]), prescriptions (RxNorm), lab tests (Logical Observation Identifiers Names and Codes [LOINC]), and clinical notes across both sites. The cohort with the highest use varied, being ICD-and-NLP at VUMC and ICD-only at MGB. Notably, despite lower overall use and a shorter study time frame (1989‐2021 at VUMC vs 1976‐2021 at MGB), VUMC cohorts had markedly more clinical notes, suggesting site-level differences in documentation practices.

OCD is a common psychiatric condition, yet remission is achieved in only one-fifth of cases [3]. The average interval between symptom onset and treatment initiation is 7 years [33], and longer duration of untreated OCD is associated with poorer outcomes [34]. Given this, identifying biological markers and genetic risk factors is crucial for facilitating earlier diagnosis and targeted interventions. Large-scale agnostic gene discovery studies of OCD are gaining traction, and recent work demonstrates that discovery yield increases as sample size grows [14]. Previous OCD GWAS have relied on cases ascertained from various sources, including specialist clinics and advocacy organizations [12,13], or national disease registers from the Nordic countries [15,16]. The availability of large-scale, real-world EHR data has provided an opportunity to develop scalable and cost-efficient tools to identify patients with psychiatric disorders, such as OCD, for genetic studies. A prior study used NLP, supplemented by structured diagnostic fields, to ascertain obsessive-compulsive symptoms (OCS) and OCD in a UK mental health case register among patients with schizophrenia, schizoaffective disorder, or bipolar disorder, with the aim of estimating prevalence and examining clinical associations [35]. However, their study was conducted at a single site within a specialty mental health setting and did not include any genetic analysis. In contrast, our study focused on identifying OCD cases from broader patient populations for genetic research, with validation across 2 US health care systems.

In this study, we developed and validated an EHR-based algorithm for identifying OCD that integrates diagnostic codes, medication records, and NLP-derived information from clinical notes. The algorithm achieved high PPV across multiple chart review rounds at 2 health care systems, VUMC and MGB, located in the southeastern and northeastern United States, respectively. We further evaluated whether algorithm-defined phenotypes captured known OCD genetic liability by testing associations with OCD PGS in individuals of European genetic ancestry from BioVU and MGBB. The significant associations observed provide early support that, within this ancestry-defined subset, the algorithm identifies cases with common variant liability overlapping that of traditionally ascertained OCD samples. As outlined below, 3 key findings from our analyses are particularly noteworthy.

First, chart reviews showed ICD-and-NLP achieved the highest PPVs (0.84 at VUMC; 0.91 at MGB), outperforming either arm alone (ICD-defined, NLP-only) and their union (ICD-or-NLP), particularly at MGB, where the PPVs for the ICD-defined and NLP-only definitions were only 0.64 and 0.51, respectively (Table 1). These findings suggest the structured diagnostic codes and NLP-derived evidence from clinical notes complement each other in enhancing the precision of OCD phenotype ascertainment. On the other hand, ICD-and-NLP identified fewer cases (515 at VUMC, 335 at MGB) for downstream genetic studies, whereas ICD-or-NLP captured more (1057 at VUMC, 1379 at MGB) but with substantially reduced PPVs (0.72 at VUMC, 0.58 at MGB). This underscores that the 2 arms are complementary only when used in conjunction since the union of their identified cases (disjunction) tends to accumulate false positives, resulting in poorer PPVs. Moreover, this pattern of findings highlights the tradeoff between achieving high precision (PPV) and maximizing the number of cases identified when selecting an algorithmic definition. Additionally, we observed PPVs achieved by using only ICD codes to select cases at the 2 US-based health care systems (PPV=0.73 and 0.64) were notably lower than those reported in the Nordic studies (0.91‐0.96 in Sweden [15], 0.85‐0.96 in Denmark [16]), suggesting less reliability of ICD-only ascertainment in the United States compared to Nordic registers.

Second, among individuals of European genetic ancestry, we observed significant associations between OCD PGS from GWAS and algorithm-defined OCD phenotypes, including ICD-defined (OR 1.22; P=1.51×10^–3 at VUMC; OR 1.30; P=7.0×10^–8 at MGB), ICD-or-NLP (OR 1.15; P=4.05×10^–3 at VUMC; OR 1.19; P=5.71×10^–7 at MGB), and ICD-and-NLP (OR 1.26; P=8.19×10^–4 at VUMC; OR 1.32; P=1.53×10^–4 at MGB). These findings provide early support that the proposed algorithm identifies phenotypes that capture common variant liability overlapping with traditionally ascertained OCD samples, supporting their potential integration with traditional cohorts to enhance the statistical power of genetic discovery. By contrast, when used in isolation, the NLP arm yielded smaller effect estimates and weaker OCD PGS associations, underscoring limitations in relying solely on rule-based NLP for ascertaining OCD cases from clinical narratives.

The MDD PGS was also associated with algorithm-defined OCD status, consistent with the high clinical comorbidity and shared genetic liability between OCD and depression. However, sensitivity analyses adjusting for MDD status showed that OCD PGS associations were more robust to this adjustment than MDD PGS associations, supporting the relative specificity of OCD-related genetic liability in the algorithm-defined OCD phenotypes. In addition, height PGS, included as a negative-control trait, was not associated with algorithm-defined OCD status at either site. These findings support the genetic plausibility and relative specificity of the algorithm-defined OCD phenotypes, while also underscoring that PGS enrichment alone should not be interpreted as definitive evidence of diagnostic specificity.

Comparing OCD PGS effect estimates across definitions, ICD-defined showed significant associations, indicating that diagnostic codes alone capture substantial common-variant liability. However, ICD-only ascertainment has lower PPV in our US EHR systems than in Nordic registers and does not capture OCD-related evidence documented only in clinical narratives. The comparison between ICD-and-NLP and ICD-or-NLP is therefore more informative for evaluating the precision-yield tradeoff. Although the stricter ICD-and-NLP definition yielded numerically higher ORs than the broader ICD-or-NLP definition at both sites, consistent with its higher PPV, confidence intervals overlapped substantially. The choice of definition should therefore be guided by the downstream application: for large-scale genetic discovery studies, where statistical power depends heavily on sample size, the broader ICD-or-NLP definition may be preferable despite modestly lower genetic enrichment; for studies prioritizing phenotypic precision, such as treatment outcome research, the stricter ICD-and-NLP definition may be more appropriate.

Finally, replication across 2 health care systems provides several recommendations for investigators seeking to apply a similar approach in other EHR systems: (1) because variability in local documentation practices and clinical workflows can affect data completeness and algorithm transferability, implementation at a new site should begin with mapping of local structured and unstructured (clinical notes) data, documentation workflows, and relevant clinical infrastructure. For example, the presence of an OCD-specialty clinic may influence how OCD-related evidence is represented in the EHR. Among patients evaluated in these settings, OCD may be more frequently documented in both structured diagnostic fields and narrative notes; at the same time, these same workflows may also generate ICD codes related to referrals or text mentions related to screening questionnaires, intake templates, or differential diagnosis. Such mentions may provide useful evidence for OCD phenotyping in some contexts but may also introduce noise for either ICD-only or NLP-only algorithms developed in other health care systems; (2) site-specific validation, including local chart reviewing, is recommended before deployment to estimate algorithm performance and identify common sources of false positives. Standardized chart review protocols and reviewer training materials (Multimedia Appendix 1) can help improve reproducibility across sites, while still allowing reviewers to account for local documentation practices; (3) the NLP component of the algorithm should be reviewed and refined against local samples before deployment, particularly for institution-specific abbreviations and acronym collisions, negation patterns, questionnaire and template formatting, and copy-forward text; and (4) when genetic data are available, PGS analyses for OCD and unrelated negative-control traits, such as height, can complement chart review by providing additional evidence that the EHR-derived phenotype captures psychiatric genetic liability overlapping with traditionally ascertained OCD samples. Together, these recommendations support an approach in which the algorithm’s overall architecture and the validation framework can be implemented across sites, while specific components, particularly the NLP rules and the chart review calibration, should be adapted to each health care system’s EHR structure, documentation practices, and clinical context.

Several limitations should be noted: (1) cross-site validation was complicated by variations in data availability and completeness. Diagnostic codes and corresponding clinical notes are not always accessible at the same time, in part because notes may be sequestered or otherwise unavailable, whereas codes remain accessible. This issue is particularly pronounced for psychiatric notes, which have historically had restricted access due to privacy safeguards. Although the ‘Open Notes’ mandate under the 21st Century Cures Act has improved access, adoption remains inconsistent across institutions. Moreover, since patients often receive care across multiple systems, only partial clinical documentation or diagnostic codes may appear within any single site’s EHR, further complicating algorithm replication; (2) rule-based NLP models, including ours, often fail to fully grasp clinical context and struggle with semantic ambiguity due to their reliance on rigid rules. Our analysis at MGB demonstrated this limitation through examples: the NLP arm of the algorithm misclassified cases due to missed negations (eg, “no” responses in poorly formatted questionnaires), copy-pasted text, and abbreviation misinterpretation (eg, “OCD” for osteochondritis dissecans); (3) PGS analyses were restricted to individuals of European genetic ancestry due to the limited availability of GWAS summary statistics for non-European populations. As a result, our PGS findings should be interpreted as supporting the genetic plausibility of the phenotype within the European-ancestry population, and not as evidence of equivalent genetic validity across other ancestral backgrounds. This limitation is not unique to this study but reflects a broader challenge in psychiatric genetics, where the underrepresentation of non-European populations in GWAS limits the cross-ancestry portability of PGS-based validation; (4) although PGS enrichment in algorithm-identified cases demonstrates elevated common variant liability for OCD, it does not establish diagnostic specificity, given genetic correlations and clinical comorbidity across psychiatric disorders. Our sensitivity analyses partially address this concern by showing that OCD PGS associations were more robust to MDD adjustment than MDD PGS associations, and that an unrelated PGS was not associated with OCD status. Nevertheless, residual confounding by psychiatric comorbidity or other clinical factors cannot be ruled out. Furthermore, although we excluded BioVU samples from the GWAS summary statistics used at VUMC and 23andMe samples at both sites, both the GWAS and our algorithm ultimately rely on clinical diagnostic conventions, and the observed PGS associations may partly reflect shared ascertainment processes rather than fully independent biological validation.

This work within the PsycheMERGE network currently focuses on developing a standardized review protocol by training chart reviewers across multiple health care institutions using simulated psychiatric patient records. This standardization and calibration aim to directly address the reduced generalizability of phenotyping algorithms due to inconsistencies in chart review strategies. Recent advances in large language models (LLMs) offer new opportunities for scalable phenotyping without substantial annotated data. Future work will explore using LLMs to identify OCD-related information in clinical notes, supported by recent findings demonstrating superior diagnostic accuracy of LLMs compared to mental health professionals in identifying OCD cases from clinical vignettes [36]. Furthermore, we will diversify the genetic ancestry representation in our OCD validation analyses by using resources such as the All of Us Research Program [37]. This will support the development of more inclusive and equitable models of OCD genetic risk, addressing the limitations of current European-centric studies.

We developed and validated an EHR-based phenotyping algorithm for OCD integrating diagnostic codes, medication records, and unstructured clinical notes across 2 health care systems (VUMC and MGB). Combining the ICD and NLP arms of this algorithm identified OCD cases with high precision, and among individuals of European genetic ancestry, the identified cases showed significant associations with GWAS-derived OCD PGS at both sites. Sensitivity analyses adjusting for MDD status and negative-control analyses using height PGS further supported the genetic plausibility and relative specificity of the phenotype. This scalable, cost-efficient approach can facilitate large-scale genetic studies and advance understanding of OCD’s complex etiology.