Study Population

We used Harvard Pilgrim Health Care (HPHC) administrative claims data to identify the initial Lyme disease cohort in Massachusetts, USA. HPHC is a not-for-profit health insurance company serving >3 million members primarily in the New England region of the United States. HPHC members are approximately half female and half male, and ≈20% of members are >65 years of age. We included HPHC members who were enrolled in medical and pharmacy benefits for >6 months from July 1, 2000, through June 30, 2019, and who were residents of Massachusetts at the time of enrollment.

To validate cases identified in the administrative claims database, we reviewed medical charts for a subset of patients with Lyme disease episodes who received care from any facility that was part of the Mass General Brigham (MGB) healthcare system. We limited chart review to a single healthcare system to simplify accessing medical charts. MGB, the largest provider system in Massachusetts, comprises 16 institutions across the care continuum and has 6,500 physicians. The system includes academic medical centers, specialty and community hospitals, and urgent and community-based care via community health centers that are geographically dispersed across eastern Massachusetts. In 2020, the MGB healthcare system was responsible for ≈20% of inpatient discharges and ≈27% of outpatient revenue in Massachusetts (12). We expected the MGB healthcare system to be representative of care delivered across the state.

Algorithm Criteria and Descriptive Analyses

Lyme disease was defined by >1 diagnosis code (ICD-9-CM code 088.81; ICD-10-CM codes A69.20, A69.21, A69.22, A69.23, and A69.29) and >1 outpatient dispensing of an antimicrobial drug used to treat Lyme disease according to Infectious Diseases Society of America guidelines (13). We defined antimicrobial drugs by using the US Food and Drug Administration National Drug Codes for doxycycline, amoxicillin, cefuroxime axetil, azithromycin, penicillin G, ceftriaxone, and cefotaxime; we included oral and nonoral formulations. We required a minimum 7-day supply of the antimicrobial drug and for it to be dispensed within 30 days of the Lyme disease diagnosis. We evaluated the use of doxycycline, amoxicillin, cefuroxime axetil, azithromycin, penicillin G, ceftriaxone, and cefotaxime to treat Lyme disease.

To identify Lyme disease episodes, we required that HPHC members did not have a Lyme disease diagnosis code documented within 180 days before meeting the Lyme disease definition (i.e., if someone had a Lyme disease diagnosis code but no antimicrobial drug dispensed and then had another Lyme disease diagnosis code <180 days later with a qualifying antimicrobial drug dispensed, we did not include the second episode). For members who had multiple Lyme disease episodes, we used recurrence intervals to exclude episodes in which the diagnosis code and antimicrobial drug were likely used for treating Lyme disease–related sequelae from the first infection; we used intervals according to those used by others for ICD-9-CM (9) and ICD-10-CM (10) codes. During the ICD-9-CM era (before October 1, 2015), the recurrence interval was 365 days. During the ICD-10 era (beginning October 1, 2015), if a member met the algorithm definition with code A69.2 (Lyme disease) or A69.20 (Lyme disease, unspecified) on or after October 1, 2015, the recurrence interval was 180 days, as long as the second Lyme disease case date was in the next calendar year. If the second Lyme disease case date was in the same calendar year, then the second episode was not included. If a member met the algorithm definition with code A69.21 (meningitis), A69.22 (other neurologic disorders), A69.23 (arthritis), or A69.29 (other conditions) on or after October 1, 2015, the recurrence interval was 365 days.

We summarized characteristics of HPHC members with algorithm-defined Lyme disease during the full study period by using descriptive statistics. We examined the frequencies and percentages of patient demographic and clinical characteristics associated with Lyme disease episodes that were available in the administrative claims data. Acute signs and symptoms were rash, fever, chills, fatigue, headache, joint and muscle pain, radiculopathy, and paresthesia, and those were identified by ICD-9-CM and ICD-10-CM diagnosis codes reported within 14 days before or after meeting the Lyme disease algorithm definition (Appendix Table). Musculoskeletal, nervous system, cardiovascular, and ocular manifestations of Lyme disease were examined up to 1 year after Lyme disease diagnosis and were also identified by diagnosis codes (Appendix Table). Among those patients with obtainable MGB medical records that were reviewed and adjudicated, we evaluated demographic and clinical characteristics and summarized acute symptoms and disseminated manifestations by using the same criteria described previously. We also assessed laboratory data captured in the medical records to determine how many cases were laboratory-confirmed.

Algorithm Validation via Medical Chart Reviews

We had an a priori goal of reviewing 125 medical charts for algorithm validation; we prioritized cases from the ICD-10-CM era and then included ICD-9-CM era episodes to obtain >125 charts. We identified 193 medical charts for persons with HPHC insurance who had evidence of Lyme disease–related care at a facility within the MGB system and who met the algorithm criteria during January 2015–June 2019; we sought medical records for a convenience sample of 171 cases.

Under the supervision of an infectious disease clinical faculty member (C.R.), 3 MGB medical residents (C.T.N., N.P., M.S.) conducted all chart abstraction and adjudication activities. Prior to conducting those activities, they received training from a Lyme disease clinical expert (J.A.). To assess interrater reliability, all 3 clinicians initially abstracted and adjudicated the same 20 medical charts. We calculated a single κ-like statistic that summarized interrater reliability across all clinicians by computing the mean of the weighted κ for each clinician pair (14). We divided the remaining charts among the 3 clinicians for single adjudications.

We conducted medical chart reviews assuming that the clinician-determined adjudication was the standard for definitively assigning Lyme disease status according to surveillance case definitions. We developed standardized abstraction and adjudication forms for chart reviews that had definitions consistent with the 2017 Council of State and Territorial Epidemiologists’ Lyme disease case definitions for confirmed, probable, and suspected cases (15) (Appendix). Abstracted data from each medical record were evidence of erythema migrans or rash; tick bite or exposure to ticks; signs and symptoms of Lyme disease; cardiovascular, musculoskeletal, or nervous system manifestations of Lyme disease; antimicrobial drugs or other medications used to treat Lyme disease; laboratory tests and results; physician diagnosis of Lyme disease; evidence of persistent signs and symptoms of Lyme disease; and healthcare facility type. Claims-based Lyme disease cases were adjudicated, and we classified each case as confirmed, probable, suspected, or not a Lyme disease case (Table 1).

We calculated positive predictive values (PPV) for claims-based Lyme disease cases adjudicated as confirmed, probable, or suspected and PPV values for confirmed or probable cases only. We calculated PPVs according to age group, healthcare facility type, period, and patients with Lyme disease–related laboratory tests to determine how performance varied across those subgroups. We used the Clopper-Pearson method to calculate 95% CIs for all PPVs (16). The study was approved by the Harvard Pilgrim Health Care Institutional Review Board.

Claims Data

From July 1, 2000, through June 30, 2019, by using the Lyme disease claims-based algorithm, we identified 12,229 Lyme disease episodes among 11,823 HPHC members who lived in Massachusetts; a total of 11,452 members had 1 Lyme disease episode, 339 had 2 episodes, and 32 had 3 or 4 qualifying episodes. Most (77.7%) episodes were identified during the ICD-9-CM era; the only applicable code was 088.81, Lyme disease. During the ICD-10-CM era, the most common cohort-defining diagnosis code was A69.20, Lyme disease unspecified (93.0%); 4.9% were identified as A69.23, arthritis due to Lyme disease; 1.4% as A69.29, other conditions associated with Lyme disease; and <1% as A69.22, other neurologic disorders in Lyme disease, or A69.21, meningitis due to Lyme disease.

We analyzed demographic and clinical characteristics of patients with Lyme disease episodes according to claims data for the overall cohort (n = 12,229) and the subset included in the chart review (n = 128) (Table 2). Most Lyme disease episodes occurred among adults >18 years of age, including 71.4% in the overall cohort (median age 42 years, interquartile range 15–55 years; 45.9% were 40–64 years of age) and 80.5% in the chart review (median age 48 years, interquartile range 29–60 years; 48.4% were 40–64 years of age). Male patients comprised 49.2% of reviewed charts and 54.6% of all Lyme disease episodes.

Of the total Lyme disease episodes, 66.3% were associated with dispensation of a >7-day supply of doxycycline, 29.4% with amoxicillin, and 4.3% with cefuroxime acetyl, azithromycin, or penicillin G. Some cases (2%) were treated with >1 antimicrobial drugs. No cases were treated with ceftriaxone or cefotaxime. Within the subset included in the chart review, the pattern was similar, although more patients (83.6%) were treated with doxycycline.

For the overall Lyme disease cohort, during the 14 days before and after the Lyme disease case date, 56.7% of cases did not have any diagnosis codes recorded in claims data that were indicative of acute signs or symptoms; 13.7% of cases had diagnosis codes for joint pain, 13.4% for rash, and 12.4% for fatigue. During the 365 days after the Lyme disease case date, 7.4% of cases had a diagnosis code indicative of a nervous system manifestation, such as Bell’s palsy, meningitis, or radiculopathy. Musculoskeletal (2.7%), ocular (2.1%), or cardiovascular (0.3%) manifestations occurred within 365 days of the Lyme disease case date, according to diagnosis codes. Those findings were generally similar among patients included in chart reviews.

Algorithm Validation via Chart Review

Of the 128 (75%) obtainable medical records that we reviewed and adjudicated, 80.5% were for cases that occurred during the ICD-10 era. The overall interrater reliability for the 20 charts reviewed by all 3 clinician adjudicators yielded a mean weighted κ of 0.94.

Overall, we adjudicated 120 of 128 reviewed charts as confirmed, probable, or suspected cases. The distribution of those 120 cases followed the expected seasonality of Lyme disease in Massachusetts; the peak was observed in July. Of the 18.8% of cases that were laboratory-confirmed (defined by positive Lyme disease culture, PCR, or standard 2-tiered tests), all were adjudicated as confirmed or probable cases. A clinical diagnosis of Lyme disease was indicated in 55.5% of charts, defined as erythema migrans or Lyme disease–associated carditis, neuroborreliosis, meningitis, or arthritis in the healthcare provider’s clinical notes. Upon chart review, erythema migrans was reported for 48% of patients (98.4% of whom were adjudicated as confirmed cases), which was substantially higher than the 15.6% of patients with evidence of rash via claims data alone. Similar to observations for claims data alone, reports of disseminated Lyme disease manifestations were uncommon upon chart review. Musculoskeletal involvement was found in 6.3%, nervous system involvement in 2.3%, cardiovascular involvement in <1%, and ocular involvement in 0% of cases; 75% (n = 9) of patients with a disseminated manifestation were adjudicated as confirmed cases.

For reviewed charts, we calculated PPVs for the algorithm overall and according to select characteristics (Table 3). Most (74.2%) charts were from patients seen in a primary care setting. The overall PPV of the algorithm for cases identified as confirmed, probable, or suspected was 93.8% (95% CI 88.1%–97.3%). When limited to confirmed or probable cases only, the PPV was 66.4% (95% CI 57.5%–74.5%). The PPV for confirmed, probable, or suspected cases was 100% (n = 25) for pediatric patients, compared with 92.2% (n = 103) for adult patients. PPVs for confirmed, probable, and suspected cases were 92.0% for those identified during the ICD-9 era and 92.4% for those identified during the ICD-10 era. When including only confirmed and probable cases, the PPV was 76.0% for the ICD-9 era and 64.1% for the ICD-10 era.

Among the 8 patients who did not have Lyme disease upon adjudication, none had erythema migrans, and 1 patient had a nonspecific rash. Only 1 patient had a documented tick bite. One patient’s chart indicated Borrelia miyamotoi infection and another noted a suspected B. miyamotoi infection. Among 5 patients who had a Lyme disease test, 4 had negative results documented.