Validation of ICD-10-CM diagnosis codes for identifying COVID-19 cases using administrative claims data in Taiwan Authors: Wei-Lun Chang, Chia-Ling Wu, Cheng-Hsun Chiu, Yu-Tung Huang Published: 2026 ABSTRACT AbstractBackgroundAccurate identification of COVID-19 cases is essential for real-world studies. This study assessed the validity of ICD-10-CM code U07.1 for identifying COVID-19 cases in administrative claims data in Taiwan.MethodsThis observational validation study linked National Health Insurance claims with laboratory-confirmed SARS-CoV-2 RT-PCR results from a multi-institutional database. We included 127,300 outpatient, emergency, and inpatient encounters from 2022. Positive and negative predictive values (PPV, NPV), sensitivity, and specificity were calculated using RT-PCR results as the reference standard, with stratified analyses conducted by clinical setting, time period, patient demographics, and hospital level.ResultsOverall, the U07.1 code demonstrated high PPV (95.3%) and specificity (99.5%), but moderate NPV (77.9%) and low sensitivity (26.6%). While PPVs remained consistently high across all clinical settings (> 91%), NPVs varied significantly, being lowest in outpatient visits (41.3%). Temporally, NPV and sensitivity were lowest during April–June 2022 (57.7% and 14.9%, respectively), coinciding with the peak outbreak and the transition toward antigen-based diagnostic policies. Sensitivity was markedly higher among inpatients (84.4%) and older adults aged ≥ 80 years (60.5%) compared to younger populations.ConclusionICD-10-CM U07.1 is a highly reliable marker for identifying confirmed COVID-19 cases in Taiwan’s claims data. However, the low overall sensitivity and substantial variability in NPV suggest significant under-capture of mild infections, particularly in outpatient settings. Researchers should exercise caution when using this code to estimate total disease burden or when defining code-negative control cohorts.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12879-026-13613-y.Keywords: COVID-19, ICD-10-CM, Administrative claims data, Validation Background Coronavirus disease 2019 (COVID-19), caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), was first identified in late 2019 and rapidly spread worldwide in early 2020. Accurate case identification for monitoring disease burden and outcomes relies on the International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM), which offers standardized diagnostic codes widely applied in claims databases and electronic health records. The specific ICD-10-CM code for COVID-19, U07.1 [COVID-19, virus identified], was implemented in the United States in April 2020 [1] and subsequently introduced in Taiwan by the National Health Insurance (NHI) Administration in May 2021 for documenting COVID-19 cases. Validation studies from several countries have assessed the performance of U07.1 in administrative data, yielding variable results. In the United States, studies of hospitalized patients reported positive predictive values (PPVs) ranging from 85% to 94% and negative predictive values (NPVs) above 95% [2–5]. In contrast, a study in Japan found poor validity when using diagnosis codes alone to identify COVID-19 hospitalizations, with a PPV of 11%. When diagnosis and procedure codes were combined, validity improved substantially, with a PPV of 71%, while NPVs consistently remained close to 99% [6]. Such variability may be attributable to differences in coding practices and case definition approaches across studies. Currently, most reports have focused on inpatient settings, with limited data from outpatient or emergency settings, and no validation studies have yet been conducted in Taiwan. In Taiwan, claims data are not routinely linked to laboratory results, and case identification relies on diagnosis codes, which may lead to misclassification and potential bias in epidemiologic studies. The objective of this study was therefore to validate the ICD-10-CM code U07.1 for identifying COVID-19 cases in Taiwan by estimating its PPV, NPV, sensitivity, and specificity, using SARS-CoV-2 reverse transcription polymerase chain reaction (RT-PCR) results as the reference standard. Methods Study design and data source This observational validation study linked Taiwan’s NHI claims data with laboratory reports from the Chang Gung Research Database (CGRD). The CGRD is the largest multi-institutional clinical electronic medical records database in Taiwan, comprising data from over 12 million patients and 10,000 beds across nine branch hospitals (Keelung, Taipei, Linkou, Taoyuan, Yunlin, Chiayi, Kaohsiung, Fengshan, and Tucheng) operated by the Chang Gung Medical Foundation. These hospitals are geographically distributed across northern, central, and southern Taiwan, and span various accreditation levels, including three medical centers, three regional hospitals, and three community hospitals—all of which are contracted with the NHI Administration [7]. Patient-level data were extracted from outpatient, emergency department, and inpatient encounters and included, but were not limited to, demographic characteristics, diagnostic and procedure codes, laboratory tests, medication prescriptions, and surgical records. Both NHI-covered and out-of-pocket services were collected to ensure the completeness of clinical information. To protect patient privacy, all data were fully de-identified before access or analysis. All procedures were conducted in accordance with the principles outlined in the Declaration of Taipei on ethical considerations regarding health databases issued by the World Medical Association. The study protocol was approved by the Institutional Review Board of the Chang Gung Medical Foundation (IRB No. 202500368B0C6001) and the informed consent was waived due to all data were fully de-identified prior to access and analysis. Study population All encounters in outpatient, emergency, or inpatient settings between January 1 and December 31, 2022, where SARS-CoV-2 RT-PCR testing was performed, were included in the study population. Encounters were excluded if RT-PCR test results were not reported on the day of or within 7 days following an outpatient or emergency department visit, or if they were not documented during the hospitalization. COVID-19 diagnosis definition COVID-19 diagnoses were identified from NHI claims records within the CGRD, encompassing outpatient, emergency department, and inpatient encounters. To ensure consistency with the NHI claims data provided by the Health and Welfare Data Science Center, Ministry of Health and Welfare—which contain only up to three diagnosis fields for outpatient and emergency encounters—analyses in these settings were restricted to the first three fields. In the NHI claims system, completion of all available diagnosis fields is not mandatory and depends on clinical documentation and coding practices. Encounters were consolidated such that multiple records for the same patient on the same day within the same clinical setting were counted as one diagnostic record, whereas visits on the same day across different settings were treated as distinct episodes. COVID-19 diagnoses were identified by the presence of the ICD-10-CM code U07.1. Three coding algorithms were evaluated: (1) any diagnosis position, defined as the presence of U07.1 in any diagnosis field (primary definition); (2) primary diagnosis only, defined as the presence of U07.1 as the first listed diagnosis; and (3) secondary or later diagnosis only, defined as the presence of U07.1 in any diagnosis field other than the first listed. Reference standard for COVID-19 cases The laboratory report with the positivity testing results from the CGRD served as the reference standard for confirming COVID-19 cases. Test results were derived from RT-PCR assays documented in outpatient, emergency department, and inpatient electronic medical records. RT-PCR testing for SARS-CoV-2 was performed on both respiratory and non-respiratory specimens, including nasopharyngeal swabs, throat swabs, saliva, sputum, and other specimens. For multiple encounters in outpatient and emergency departments with RT-PCR testing on the same day, any positive result was counted as positive. Positivity was determined based on the earliest available RT-PCR result on the encounter date or within 7 days thereafter. For inpatients, all RT-PCR results obtained during hospitalization were reviewed, and the encounter was classified as positive if any result was positive. Statistical analysis PPV, NPV, sensitivity, and specificity, with corresponding 95% confidence intervals (CIs), were estimated using RT-PCR test results as the reference standard. Predictive values (PPVs and NPVs) were calculated overall for each of the three coding algorithms. For the any-position diagnosis definition, which served as the primary definition, PPV, NPV, sensitivity, and specificity were further evaluated in stratified analyses by clinical setting (outpatient, emergency, and inpatient), time period (January–March, April–June, July–September, and October–December 2022), sex (male and female), age group (< 20, 20–39, 40–59, 60–79, and ≥ 80 years), and hospital level (medical center and non-medical center). Additionally, a sensitivity analysis was conducted to assess the robustness of the predictive values. Outpatient and emergency encounters were evaluated using up to six diagnosis fields, as available in the CGRD, rather than restricting to the first three fields, while inpatient encounters retained the same number of diagnosis fields as in the main analysis. All statistical analyses were performed using SAS software, version 9.4 (SAS Institute Inc., Cary, NC, USA). Results Characteristics of study population A total of 127,300 encounters with SARS-CoV-2 RT-PCR testing conducted between January 1 and December 31, 2022, were included in the study. The extraction process of study subjects is shown in Fig. 1. Most encounters occurred in the emergency department (87,152; 68.5%), followed by outpatient (24,080; 18.9%) and inpatient settings (16,068; 12.6%). Nearly half of the encounters (55,817; 43.8%) took place between April and June 2022, when the major outbreak began in Taiwan. More than half of the patients were male (67,677; 53.2%), whereas female patients accounted for 46.8%. The major age group was 60–79 years (39,073; 30.7%), followed by 40–59 years (29,265; 23.0%) and < 20 years (21,810; 17.1%). In addition, over half of the encounters occurred in medical centers (67,113; 52.7%), while 47.3% (60,187) were in non-medical centers (Table 1). Flowchart of study population selection. Note: RT-PCR, Reverse transcription polymerase chain reaction Table 1Characteristics of patient encounters included in the studyCharacteristicTotalRT-PCR (+)RT-PCR (-) n (%) n (%) n (%) Overall 127,30035,430(27.8)91,870(72.2) Clinical setting Outpatient24,080(18.9)14,952(42.2)9,128(9.9) Emergency87,152(68.5)15,016(42.4)72,136(78.5) Inpatient16,068(12.6)5,462(15.4)10,606(11.5) Time period January–March 202231,568(24.8)58(0.2)31,510(34.3) April–June 202255,817(43.8)25,730(72.6)30,087(32.7) July–September 202232,399(25.5)6,479(18.3)25,920(28.2) October–December 20227,516(5.9)3,163(8.9)4,353(4.7) Sex Male67,677(53.2)17,625(49.7)50,052(54.5) Female59,623(46.8)17,805(50.3)41,818(45.5)Age group, years <2021,810(17.1)8,905(25.1)12,905(14.0) 20–3919,186(15.1)7,382(20.8)11,804(12.8) 40–5929,265(23.0)7,858(22.2)21,407(23.3) 60–7939,073(30.7)7,727(21.8)31,346(34.1) ≥8017,966(14.1)3,558(10.0)14,408(15.7) Hospital level Medical center67,113(52.7)15,796(44.6)51,317(55.9) Non-medical center60,187(47.3)19,634(55.4)40,553(44.1)Note: RT-PCR, Reverse Transcription Polymerase Chain Reaction Characteristics of patient encounters included in the study Note: RT-PCR, Reverse Transcription Polymerase Chain Reaction Of all encounters, 35,430 (27.8%) were RT-PCR positive and 91,870 (72.2%) were negative. Positive cases were more common in outpatient settings, peaked in April–June 2022, and occurred more frequently in younger patients, whereas negative cases were more frequent in medical centers and among older patients (Table 1). Validity The PPV of the ICD-10-CM code U07.1 for COVID-19 was consistently high across the three coding algorithms. The PPV was 95.3% (95% CI, 94.9%–95.8%), 97.3% (95% CI, 96.9%–97.7%), and 93.1% (95% CI, 92.4%–93.9%) for “any diagnosis position”, “primary diagnosis only”, and “secondary or later diagnoses only,” respectively (Table 2). In “any diagnosis position” definition, PPVs were highest in the emergency department (99.4%) and outpatient settings (98.1%), while the inpatient setting showed a lower PPV of 91.9%. When stratified by time period, the PPV was markedly lower in January–March 2022 (63.0%), whereas the PPVs in the later quarters were all above 94% (94.6% in April–June, 94.8% in July–September, and 98.3% in October–December). PPVs were similar by sex, with marginally higher values in females (95.7%) than males (95.1%). All PPVs remained consistently high across age groups, ranging from 93.2% in those aged 40–59 years to 98.2% in those aged < 20 years. Similarly, PPVs were 95.2% in medical centers and 95.7% in non-medical centers (Fig. 2). Table 2Predictive values of ICD-10-CM code U07.1 under three coding algorithmsCoding algorithmPositive predictive valueNegative predictive valueRT-PCR (+)& DiagnosisDiagnosis% (95% CI)RT-PCR (-)& Non-diagnosisNon-diagnosis% (95% CI)Any position (primary definition)9,4289,88895.3 (94.9–95.8)91,410117,41277.9 (77.6–78.1)Primary diagnosis only5,2595,40597.3 (96.9–97.7)91,724121,89575.2 (75.0-75.5)Secondary or later diagnosis only4,4464,77393.1 (92.4–93.9)91,543122,52774.7 (74.5–75.0)Note: CI, Confidence interval Predictive values of ICD-10-CM code U07.1 under three coding algorithms Note: CI, Confidence interval Overall and stratified positive predictive value of the ICD-10-CM code U07.1 for COVID-19. Note: PPV, positive predictive value; CI, confidence interval The NPV of the ICD-10-CM code U07.1 for COVID-19 was moderate across the three coding algorithms. The NPV was 77.9% (95% CI, 77.6%–78.1%), 75.2% (95% CI, 75.0%–75.5%), and 74.7% (95% CI, 74.5%–75.0%) for “any diagnosis position”, “primary diagnosis only”, and “secondary or later diagnoses only,” respectively (Table 2). In “any diagnosis position” definition, the NPV was highest in inpatient encounters (92.3%) and emergency department visits (85.5%), whereas outpatient encounters showed a markedly lower NPV of 41.3%. When stratified by time period, the NPV was nearly perfect in January–March 2022 (99.9%; 95% CI, 99.9–100.0), but declined in subsequent quarters: 57.7% in April–June, 89.2% in July–September, and 82.1% in October–December. Male patients had a higher NPV (80.0%) compared with female patients (75.4%). The NPV increased with age, ranging from 64.1% among patients aged < 20 years to 91.1% in those aged ≥ 80 years. The NPVs were higher in medical centers (84.9%) than in non-medical centers (70.5%) (Fig. 3). Overall and stratified negative predictive value of the ICD-10-CM code U07.1 for COVID-19. Note: NPV, negative predictive value; CI, confidence interval The sensitivity of the ICD-10-CM code U07.1 for identifying RT-PCR–confirmed COVID-19 cases were overall low at 26.6% (95% CI, 26.2%–27.1%). The highest value was observed in inpatients (84.4%), compared to 18.4% in the emergency department and 13.7% in outpatient settings. Sensitivity increased over time, from 14.9% in April–June 2022 to 70.3% in October–December, and was higher among older patients (60.5% in those aged ≥ 80 years and 42.5% in those aged 60–79 years) compared to younger groups (< 20 years: 19.1%; 20–39 years: 9.8%). It was slightly higher in males (29.4%) than in females (23.8%) and was greater in medical centers (42.5%) than in non-medical centers (13.8%) (Fig. 4). In contrast, specificity was consistently high across all subgroups, with an overall value of 99.5% (95% CI, 99.5%–99.5%). Specificity exceeded 99% across all clinical settings, time periods, sexes, age groups, and hospital levels, except among inpatients, where it was slightly lower at 96.2% (Fig. 5). Overall and stratified sensitivity of the ICD-10-CM code U07.1 for COVID-19. Note: CI, confidence interval Overall and stratified specificity of the ICD-10-CM code U07.1 for COVID-19. Note: CI, confidence interval Sensitivity analysis In the sensitivity analysis using up to six diagnosis fields for outpatient and emergency encounters, the PPV (95.4%) and NPV (78.0%) were nearly identical to those under the primary definition with three fields (PPV 95.3%, NPV 77.9%), indicating that expanding the number of diagnosis fields did not materially affect predictive values (Supplementary Table 1). Discussion This large validation study in Taiwan demonstrated that the ICD-10-CM code U07.1 exhibited a consistently high PPV. In contrast, the NPV was moderate and varied across clinical settings, time periods, and age groups. Specificity remained consistently high, while sensitivity was comparatively lower, particularly in certain subgroups. These findings support the reliable use of U07.1 for identifying true cases, while highlighting the need for caution when using this code to estimate disease burden or when defining code-negative populations in claims-based research. The consistently high PPV observed in our study aligns with previous U.S. inpatient validation studies, which also reported PPVs exceeding 90% [2, 3, 5, 8]. In contrast, Ishiguro et al. (2024) [6] reported lower validity in Japan when diagnosis codes were used alone, with improved accuracy observed only when combined with procedure codes. Our overall NPV estimates were slightly lower than those reported in prior inpatient studies [2, 3, 8], likely because previous work was restricted to hospitalized populations, whereas our study also included outpatient and emergency settings, where coding completeness may differ. In particular, NPV was lowest in the outpatient setting compared with emergency and inpatient settings. This likely reflects differences in clinical presentation, as outpatient encounters are more likely to involve mild or asymptomatic cases [9], which are less likely to be assigned a COVID-19 diagnosis. A similar pattern was observed among younger patients, who are also more likely to present with mild or asymptomatic disease [10], resulting in lower coding completeness and reduced NPV compared with older patients, who often present with more pronounced symptoms or comorbidities [11]. This under-capture of mild infections leads to a higher proportion of false negatives and consequently a lower NPV, and the same mechanism also contributes to lower sensitivity in these subgroups. Consistent with our findings, Bhatt et al. (2021) [2] also reported age-related variation in sensitivity, with higher sensitivity among older patients. Temporal trends were also evident. Because NPV is influenced by disease prevalence, temporal variations in NPV may partly reflect changes in underlying infection prevalence during the study period. The lower NPV during April to June 2022 likely reflects the early phase of the COVID-19 outbreak in Taiwan [12], when case numbers surged rapidly and diagnostic as well as administrative workflows had not yet fully stabilized. In the latter half of 2022, NPV improved as operational processes became more structured and linkages between laboratory results and diagnostic coding systems were strengthened, reducing the number of false negatives. These early operational challenges likely contributed to both lower NPV and reduced sensitivity. Similar temporal improvements in coding performance have been reported elsewhere, as healthcare systems adapted to the evolving pandemic and coding practices became more standardized [2]. Furthermore, the temporal fluctuations in validity metrics, particularly the decline in NPV and sensitivity during the second quarter of 2022, should be interpreted in the context of Taiwan’s evolving diagnostic policies. During the major outbreak from April to June 2022, the Taiwan government adopted a policy in late May that recognized a positive rapid antigen test confirmed by a physician as a COVID-19 case. Because our study utilized RT-PCR as the sole reference standard, this policy shift likely contributed to the “under-capture” of cases in our PCR-verified cohort. Many patients diagnosed via antigen tests during the surge might not have undergone concurrent RT-PCR testing, leading to a higher proportion of cases being coded as U07.1 without a corresponding positive RT-PCR record in the database. This administrative and diagnostic transition, coupled with the immense pressure on healthcare workflows during the peak of the pandemic, likely explains the observed dip in NPV and sensitivity during this specific period. To investigate the observed low sensitivity, we examined diagnosis codes among PCR-positive encounters without U07.1 using all available diagnosis fields and identified three patterns. First, some encounters occurred in administrative or screening contexts (e.g., pre-procedural screening), including Z04.8, Z11.52, and Z01.818. Second, clinically compatible presentations were recorded, including upper respiratory tract infection (J06.9), fever (R50.9), and pneumonia (J18.9). Third, non-specific alternative codes were also observed, including B97.89 and B34.8. These findings suggest that reliance on U07.1 alone may underestimate infection burden in claims data. Future studies should consider broader case definitions and conduct sensitivity analyses with alternative outcome definitions to assess the robustness of the findings. Taken together, these findings have important implications for claims-based COVID-19 research. While the high PPV and specificity of U07.1 support its use for case identification, its low sensitivity indicates under-ascertainment of infections. Consistently, the moderate NPV suggests that code-negative individuals may include undetected infections, leading to outcome misclassification and potentially affecting the validity of comparative analyses. This study has some limitations. First, although RT-PCR served as the reference standard, it is not perfect; its accuracy is strongly influenced by the timing of testing, with early or late sampling more likely to yield false-negative results because viral load may be below the detection threshold [13]. Such misclassification may lead to an underestimation of sensitivity and bias in NPV estimates. Nevertheless, RT-PCR remained the most accurate diagnostic tool available during the study period, and testing protocols in Taiwan were highly standardized, making the impact of this limitation minimal and unlikely to alter the main conclusions. Second, antigen test results were not included. Since rapid antigen tests were only accepted as diagnostic confirmation in Taiwan after May 2022 and their implementation was inconsistent across settings, using RT-PCR as the reference standard, although imperfect, provided a more standardized and reliable basis for evaluating diagnostic code performance. Third, the CGRD-derived population may overrepresent patients with more severe conditions [7], which may limit generalizability to all individuals with COVID-19; however, its generalizability to patients seeking medical care is likely less affected. Conclusions The ICD-10-CM code U07.1 demonstrated consistently high PPV and specificity in Taiwan, supporting its reliability for case identification in claims-based research. However, caution is warranted when interpreting code-negative populations given the variability of NPV and sensitivity. Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1: Supplementary Table 1. Sensitivity analysis of the ICD-10-CM code U07.1 for COVID-19, based on outpatient and emergency encounters with six diagnosis fields Supplementary Table 1 Supplementary Table 1 Acknowledgements The authors acknowledge the Center for Big Data Analytics and Statistics at Chang Gung Memorial Hospital, Linkou, Taiwan, for assistance with statistical and data analysis. Abbreviations Chang Gung Research Database Confidence interval Coronavirus disease 2019 International Classification of Diseases, Tenth Revision, Clinical Modification National Health Insurance Negative predictive value Positive predictive value Reverse transcription polymerase chain reaction Severe acute respiratory syndrome coronavirus 2 Author contributions WLC developed the study protocol under the supervision of CHC and YTH. WLC performed the statistical analyses, and CLW prepared the figures and tables. WLC and YTH drafted the initial manuscript, and all authors revised, read, and approved the final version. Funding This study was financially supported by the Chang Gung Medical Foundation, Taiwan (Project No. CGRPG3Q0021), which had no role in the study design, data collection, analysis, interpretation of the results, or the writing of the manuscript. Data availability Data cannot be shared publicly due to the Chang Gung Medical Foundation’s policies, which are based on the Personal Data Protection Act. All study data have been encrypted to address privacy concerns, yet they remain linkable for research purposes and are limited to use exclusively at the Chang Gung Memorial Hospital. Declarations Ethics approval and consent to participate The study protocol was approved by the Institutional Review Board of the Chang Gung Medical Foundation (IRB No. 202500368B0C6001). All data were fully de-identified prior to access and analysis; therefore, the requirement for informed consent was waived. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1: Supplementary Table 1. Sensitivity analysis of the ICD-10-CM code U07.1 for COVID-19, based on outpatient and emergency encounters with six diagnosis fields Supplementary Table 1 Supplementary Table 1 Data Availability Statement Data cannot be shared publicly due to the Chang Gung Medical Foundation’s policies, which are based on the Personal Data Protection Act. All study data have been encrypted to address privacy concerns, yet they remain linkable for research purposes and are limited to use exclusively at the Chang Gung Memorial Hospital.