Safety, treatment delivery, and pathological response before and after pembrolizumab implementation for early ER-negative, HER2-normal breast cancer Authors: AF Henriksen, ANH Lundgren, MA Andersen, TA Gerds, AS Knop, TS Petersen, U Lassen, A Bjerrum Published: 2026 ABSTRACT AbstractBackgroundPembrolizumab has become standard neoadjuvant systemic therapy (NST) for early estrogen receptor (ER)-negative, human epidermal growth factor receptor 2 (HER2)-normal early breast cancer (eBC). Electronic health record (EHR) data may support real-world assessment of its implementation in routine care.Patients and methodsUsing centralized EHR data from eastern Denmark, we identified adults with ER-negative, HER2-normal eBC receiving NST. Patients were grouped by treatment start before (April 2021-June 2022; comparator) or after (April 2023-June 2024; study) pembrolizumab implementation. We compared treatment exposure, adverse events (AEs) from laboratory data, clinical notes, and AE-indicating medications, and pathological complete response (pCR) from pathology records. Follow-up ended 200 days after surgery, progression, or death. Manual validation included 50 patients.ResultsAmong 210 included patients, 109 were in the study population and 101 in the comparator population. Baseline characteristics were broadly similar. Pembrolizumab was received by 77% versus 0% and carboplatin omission was less frequent after implementation (7% versus 47%). During NST, grade 3/4 AEs were observed in 77 study patients (71%) and 66 comparator patients (65%). Suspected grade 3/4 autoimmune disorders during NST occurred in 18 pembrolizumab-treated study patients (21%) and in no comparator patients, and no grade 5 AEs were identified. pCR was achieved in 74 study patients (68%) versus 49 comparator patients (49%). Manual validation showed high structured-data completeness after resolution of a small number of data processing errors but also identified missing data.ConclusionAfter introduction of pembrolizumab into NST, patients with early ER-negative, HER2-normal breast cancer (BC) had a higher pCR proportion, alongside a higher toxicity.Key words: ER-negative HER2-normal breast cancer, neoadjuvant systemic therapy, real-world evidence Introduction Higher rates of pathological complete response (pCR) after neoadjuvant systemic therapy (NST) have been associated with both better event-free survival and overall survival (OS) in early breast cancer (eBC),1 and are used as a surrogate marker for long-term clinical outcomes in the treatment of estrogen receptor-negative (ER-negative), human epidermal growth factor receptor 2 normal (HER2-normal) eBC. Based on a decision in April 2023 by the Danish Medicines Council,2 current Danish national clinical guidelines recommend a regimen including epirubicin and cyclophosphamide (EC), paclitaxel, carboplatin, and pembrolizumab as NST in ER-negative, HER2-normal eBC.3 EC can be prescribed every 2 weeks (q2w) as dose-dense (dd) or every 3 weeks (q3w) at an identical cumulative dose. The addition of the programmed cell death protein 1 checkpoint inhibitor pembrolizumab to standard of care (SoC) was based on the results from the KEYNOTE-522 trial,4,5 where both improvement of pCR rates and OS were seen in the pembrolizumab arm, accompanied by more frequent adverse events (AEs). Immune-related AEs (irAEs) in particular warrant close attention, as they may be severe and potentially long-lasting. Because external validity of randomized controlled trial results might vary depending on study population selection, complementary real-world data (RWD) can be utilized to generate real-world evidence (RWE) on complete patient cohorts as seen in the clinic. Utilizing RWE aligns with the wishes of the Danish Medicines Council, who advised that data be collected to monitor both pCR outcomes and safety postapproval of pembrolizumab.2 Here, we present results from analyses of electronic health record (EHR) data, comparing safety and effectiveness outcomes of NST for ER-negative, HER2-normal eBC treatment in two periods: before (comparator population, SoC: q2w EC followed by carboplatin + paclitaxel) and after (study population, SoC: pembrolizumab alongside q2w EC followed by carboplatin + paclitaxel or pembrolizumab alongside carboplatin + paclitaxel followed by q3w EC) introducing pembrolizumab to SoC for ER-negative, HER2-normal eBC in Denmark. The primary objective was to describe and compare AEs captured through structured data (laboratory results), unstructured EHR text (clinical notes), and prescription patterns of AE-indicating medications before, during, and after NST in the two populations. The secondary objective was to describe proportions of pCR after NST treatment in the two populations. Materials and methods Eligibility criteria for study population and comparator population We screened all adult patients who resided in Region Zealand or the Capital Region of Denmark and had an International Classification of Diseases-10 diagnosis code for breast cancer assigned between 1 January 2016 and 30 June 2024 and who had at least one hospital contact in an oncology department after the diagnosis code assignment. We then identified patients who received at least one NST dose (outside clinical trials) in the Capital Region of Denmark for breast cancer between April 2021 and June 2024 and who had the ER-negative, HER2-normal pathological subtype. This cohort was divided into the study population (treatment initiation between 1 April 2023 and 30 June 2024) and the comparator population, defined over a temporally shifted, matched treatment initiation window (1 April 2021 to 30 June 2022). The study population eligibility period reflects the approval of pembrolizumab as SoC in Denmark (April 2023), with the end date being the approved inclusion cut-off date. The comparator window was selected to mirror the duration and calendar months of the study period, providing a similarly sized preimplementation cohort while keeping manual free-text AE review feasible. Data sources and data collection EHR data were extracted on 11 April 2025 from a centralized database covering all public hospitals treating breast cancer in eastern Denmark. Pathology records were extracted on 3 February 2025, from the Danish national pathology database. Extracted data covered EHR records from January 2016 to March 2025 and pathology records from January 1978 to January 2025. Selected data from February to December 2025 were manually extracted. Data included structured records on antineoplastic treatments, selected supportive medications, treatment plan titles, laboratory results, and pathology records. Manual review of free-text clinical notes was used to ascertain menopause status and narratively documented AEs not captured in structured-data fields. Laboratory-defined AEs were obtained from structured laboratory data. Eastern Cooperative Oncology Group (ECOG) performance status was extracted by both manual review and a custom rule-based R function. Cause of death in deceased patients and possible cancer stage I misclassification were assessed by manual review. See Supplementary Materials (available at https://doi.org/10.1016/j.esmorw.2026.100769) for further details. Index date and observation periods Index date was the first NST treatment day. The observation period started 90 days before the index date for manual record review. For baseline medications and pathology records, the observation period started 730 days before the index date and 1 January 1978, respectively. We followed patients for 200 days after eBC surgery or until cancer progression (documented in free-text notes) or death, whichever occurred first. Definitions Baseline variables were ascertained from the most recent recorded measurement within 90 days before the index. Medication exposures were summarized from anatomical therapeutic chemical-coded prescriptions and administrations across prespecified pre-, peri-, and post-NST periods (from 730 days before the index to 200 days after breast cancer operation). ER-negative was defined as ER 0% or ER 1%-9% with a nonluminal PAM50 subtype. In descriptive tables, ER status was defined as ER-low (ER 1%-9%) versus ER 0. HER2-normal was defined as immunohistochemistry (IHC) 0, 1+, or 2+ with negative in situ hybridization (ISH). In descriptive tables, HER2 status was defined as HER2-low (IHC 1+ or 2+ with negative ISH) versus HER2 0/ultralow. Modified Charlson Comorbidity Index excluding malignant diagnoses was reported. Clinical T and N stages were obtained from the most recently signed clinical cancer-staging record within 90 days before to 200 days after index, allowing for delayed structured registration of baseline stage. Antineoplastic treatment exposure was summarized by drug relative to the Danish standard neoadjuvant regimen, with doses standardized to mg/m2 or carboplatin area under the concentration-time curve. Drug-specific measures of treatment duration, dose cessations, dose reductions, relative cumulated dose (RCD), relative treatment time (RTT), dose intensity (DI), and relative dose intensity (RDI) were calculated from administered drug data. The EC values were calculated relative to the q3w regimen. See Supplementary Materials (available at https://doi.org/10.1016/j.esmorw.2026.100769) for details. AEs were recorded throughout follow-up, including the adjuvant period where applicable. Events were graded according to National Cancer Institute Common Terminology Criteria for Adverse Events (NCI CTCAE), version 5.0,6 and summarized as grades 1-2 and 3-4. For recurrent events, the highest grade per treatment cycle was used. Overlapping unstructured AE terms were consolidated into clinically coherent groupings based on shared or overlapping symptomatology and the highest grade within each group was registered. AE groups with a frequency of ≤5 patients and limited to grade 1-2 severity were not reported. Suspected autoimmune AEs were assigned a CTCAE-defined label according to the recorded organ-specific manifestation. A presumed immune-mediated label was added when the treating physician’s narrative note described the event as suspected autoimmune or immune-related toxicity. No independent adjudication was carried out. Available routine laboratory data included irAE-relevant tests such as thyroid-stimulating hormone, troponin, and cortisol as ordered by the treating physician. Laboratory markers without a clear CTCAE v5.0 grading cut-off were not tabulated as structured AE outcomes; for troponin, any value above the upper limit of normal was described instead of being assigned a CTCAE grade. Selected supportive medications, including drugs recommended in regional guidelines for treatment of suspected immune-related toxicity, were summarized as indirect indicators of toxicity management and used to supplement laboratory- and free-text-based AE assessment. pCR and tumor infiltrating lymphocytes (TILs) were defined according to Danish national clinical guidelines,7 and both were extracted from breast cancer operation pathology records identified between the index date and 250 days after the index date. Full definitions are provided in the Supplementary Materials (available at https://doi.org/10.1016/j.esmorw.2026.100769). Missing and zero values For baseline characteristics, percentage missing values were calculated for all variables, except for the Charlson index, in which the absence of a diagnosis code was interpreted as the absence of comorbidity. Structured laboratory AEs were summarized among patients with available values. Data validation To verify concordance between values displayed in the EHR clinical interface and those extracted from the underlying source database, pragmatic targeted manual validation was carried out by AFH in 50 patients. Across 98 predefined variables and study-specific time windows, values obtained via manual chart review were compared with those derived from the source database; any discrepancy was classified as an error, regardless of its presumed origin. Each error was reviewed by AFH through cross-referencing of the clinical EHR interface, extracted source data, and derived analysis variables to identify the most likely explanation. A sample size of 50 observations per variable was chosen to achieve a high probability of detecting errors if the true error rate was ≥5% (92% probability of at least one observed error). In the absence of observed errors, the one-sided 95% upper confidence bound on the error rate is 5.8%. This sample size was chosen as a pragmatic variable-level error detection to identify potential issues across selected key variables. Validated variables included key demographic information, staging, laboratory results, BC subtype, treatments, surgery, and supportive medication variables. The custom R-based ECOG performance status extraction was also informally compared with manually abstracted values, with discordances adjudicated by AFH before analysis. See Supplementary Materials (available at https://doi.org/10.1016/j.esmorw.2026.100769) for details. Discrepancies identified during validation were corrected before final analyses. Statistics and software All statistical analyses were conducted using R8 (R Foundation for Statistical Computing, Vienna, Austria) in RStudio within Microsoft Azure Machine Learning Studio, in which pseudonymized data were stored. Baseline covariates were compared using Welch’s two-sample t test for continuous variables and Fisher’s exact test for categorical variables, applied to nonmissing values. Normality of body mass index and age was assessed by visual inspection of histograms, density curves, Q–Q plots, and boxplots. Variables satisfying the normality assumption are reported as mean (standard deviation), and nonnormal variables as median (interquartile range). Results Figure 1 shows eligibility criteria. Two patients in the comparator population were excluded due to other reasons and 27 patients in total were excluded due to being enrolled in clinical trials. Eligibility flow chart. ATC, anatomical therapeutic chemical; BC, breast cancer; ER−, estrogen receptor-negative; EPIC, Epic Systems Corporation; HER2-normal, human epidermal growth factor receptor 2 normal; ICD-10, International Classification of Diseases, 10th Revision; NST, neoadjuvant systemic therapy. Baseline/index characteristics Table 1 shows baseline characteristics before the NST starts. Supplementary Table S1, available at https://doi.org/10.1016/j.esmorw.2026.100769, shows baseline CTCAE values for structured AEs, with grade 0 retained to indicate available measurements without CTCAE-defined abnormality. Table 1Patient and tumor baseline characteristicsCharacteristicComparator populationStudy populationP valuea(N = 101)Missing(N = 109)MissingSex——Not testable Female101 (100%)109 (100%)Age——0.93 Median (IQR)54 (43, 68)56 (42, 66)Menopausal status—3%0.89 Postmenopausal54 (53%)58 (55%) Premenopausal47 (47%)48 (45%)BMI——0.51 Mean (SD)26.6 (5.1)27.1 (5.5)ECOG performance status15%10%0.79 080 (93%)89 (91%) 16 (7%)9 (9%) 20 (0%)0 (0%)Modified Charlson Comorbidity Index——0.17 096 (95%)96 (88%) 14 (4%)11 (10%) ≥21 (1%)2 (2%)ER status——1.00 Low2 (2%)2 (2%) 099 (98%)107 (98%)HER2 status——0.01 Low99 (98%)96 (88%) 0/ultralow2 (2%)13 (12%)N stage8%5%0.45 N066 (71%)68 (65%) N+27 (29%)36 (35%)Stage9%5%0.56 I11 (12%)8 (8%) II72 (78%)83 (80%) III9 (10%)13 (12%)Percentages calculated for non-NA values.BMI, body mass index; ECOG, Eastern Cooperative Oncology Group; ER, estrogen receptor; HER2, human epidermal growth factor receptor 2; IQR, interquartile range; NA, not applicable; SD, standard deviation.aWelch two-sample t test (BMI and age); two-sided Fisher’s exact test (categorical)—both on non-NA values. Patient and tumor baseline characteristics Percentages calculated for non-NA values. BMI, body mass index; ECOG, Eastern Cooperative Oncology Group; ER, estrogen receptor; HER2, human epidermal growth factor receptor 2; IQR, interquartile range; NA, not applicable; SD, standard deviation. Welch two-sample t test (BMI and age); two-sided Fisher’s exact test (categorical)—both on non-NA values. All included patients were female and approximately half were postmenopausal. In the comparator population, a higher proportion had low HER2 receptor expression. No other significant differences between the two groups were seen, although the different number of missing values for ECOG performance status, cancer stage, and hypoalbuminemia should be noted. Nineteen patients were recorded as having stage I disease before the index date and were therefore outside the recommended target population for NST; most had tumor sizes close to 2 cm. Antineoplastic treatment characteristics Figure 2 shows summarized NST treatment characteristics for paclitaxel, carboplatin, epirubicin, and cyclophosphamide in the two populations as overlapping violin-, scatter-, and boxplots, with detailed numerical results provided in Supplementary Table S2, available at https://doi.org/10.1016/j.esmorw.2026.100769. Figure 3 is a swimmer’s plot showing order and length of NST subregimens (epirubicin–cyclophosphamide versus carboplatin–paclitaxel) with indicators for whether a patient received any pembrolizumab and/or carboplatin as part of the treatment and if EC were intended as q2w. Scatterplot, boxplot, and violin plot for RTT, RCD, and RDI by drug and population.Y-axes represent relative value in relation to recommended standard drug courses (dashed line) of four administrations (12 for paclitaxel) over 84 days. Y-axes are lower bounded at 0. RCD, relative cumulated dose; RDI, relative dose intensity; RTT, relative treatment time. Swimmer’s plot of neoadjuvant systemic therapy subregimens by population. Ordered by first subregimen and then length of combined treatment. Black outlines indicate courses with intercalated regimens. Treatment start is the first neoadjuvant treatment administration. Treatment stop is the earliest of the progression date, death date, or surgery date minus 7 (minus 14 for paclitaxel and q2w EC). EC, epirubicin and cyclophosphamide. Pembrolizumab was administered to 84 patients (77%) in the study population and to none in the comparator population. The RDI for cyclophosphamide and epirubicin were comparable between the two populations and aligned closely with the q3w regimen standard, despite approximately half of patients being prescribed q2w EC. Paclitaxel and carboplatin both exhibited higher RDI in the study population, primarily driven by a higher RCD of both drugs. A marked difference in the proportion of patients not receiving any dose of carboplatin was observed (7% in the study population versus 47% in the comparator population). In the study population, 36 patients (33%) received at least one dose of adjuvant therapy versus 44 patients (44%) in the comparator population. See Supplementary Materials (available at https://doi.org/10.1016/j.esmorw.2026.100769) for further details. AEs Table 2 shows unstructured AEs and selected structured AEs for the NST period; Supplementary Table S3, available at https://doi.org/10.1016/j.esmorw.2026.100769, shows AEs for the complete follow-up period. For structured AEs during NST, no values were missing for anemia, neutropenia, thrombocytopenia, or creatinine; missingness for alanine transaminase, bilirubin, and hypoalbuminemia was 1% versus 2%, 1% versus 3%, and 3% versus 20% in the study and comparator populations, respectively. Troponin measurements were available for 71% of the study population and 4% of the comparator population; among tested patients, elevated values occurred in 29% and 0%, respectively. Table 2AE frequencies in the neoadjuvant systemic therapy periodAdverse eventComparator population (N = 101)Study population (N = 109)Any gradeGrade 3/4Any gradeGrade 3/4Anemiaa87 (86%)8 (8%)107 (98%)26 (24%)Fatigue97 (96%)1 (1%)106 (97%)4 (4%)Nausea/vomiting88 (87%)90 (83%)Pain (including)76 (75%)1 (1%)87 (80%)4 (4%) Headache28 (28%)1 (1%)31 (28%) Abdominal pain20 (20%)25 (23%)1 (1%)Peripheral sensory neuropathy78 (77%)87 (80%)1 (1%)Mucositis oral/oropharyngeal pain/esophagitis/pharyngitis69 (68%)1 (1%)84 (77%)Constipation60 (59%)72 (66%)1 (1%)Neutropeniaa73 (72%)45 (45%)68 (62%)43 (39%)Diarrhea42 (42%)1 (1%)65 (60%)1 (1%)Alanine transaminase increaseda37 (37%)1 (1%)64 (59%)4 (4%)Hypoalbuminemiaa25 (31%)62 (58%)1 (1%)Thrombocytopeniaa42 (42%)9 (9%)60 (55%)9 (8%)Alopecia43 (43%)55 (50%)Anorexia/weight loss32 (32%)48 (44%)Eczema/erythema/rash/pruritus/skin-nail infection30 (30%)48 (44%)3 (3%)Syncope/presyncope/dizziness34 (34%)2 (2%)43 (39%)2 (2%)Flu-like symptoms/chills30 (30%)40 (37%)Fever36 (36%)1 (1%)39 (36%)Dry eye/mouth/mucosal membranes33 (33%)36 (33%)Dyspnea22 (22%)36 (33%)Hot flashes29 (29%)35 (32%)Cough/lung infection/pneumonitis24 (24%)29 (27%)2 (2%)Muscle weakness16 (16%)28 (26%)1 (1%)Dyspepsia19 (19%)27 (25%)Edema16 (16%)24 (22%)Eye/ear/nose/sinus/tooth inflammation-infection11 (11%)22 (20%)Symptoms of anxiety/depression/cognitive disturbance12 (12%)20 (18%)Allergic reaction/infusion reaction15 (15%)19 (17%)Creatinine increaseda4 (4%)18 (17%)3 (3%)Palpitations13 (13%)17 (16%)Dysuria/hematuria/genitourinary infection8 (8%)16 (15%)1 (1%)Palmar-plantar erythrodysesthesia syndrome5 (5%)13 (12%)Dysgeusia/anosmia6 (6%)12 (11%)Upper respiratory infection13 (13%)12 (11%)Colitis/lower gastrointestinal hemorrhage3 (3%)11 (10%)1 (1%)Blurred vision/flashing lights/photophobia6 (6%)10 (9%)Hemorrhoids/anal fistula/fissure6 (6%)10 (9%)Epistaxis/hematoma8 (8%)8 (7%)Arthritis1 (1%)5 (5%)Hypotension4 (4%)3 (3%)1 (1%)Hypertension1 (1%)1 (1%)1 (1%)Arranged by AE frequency in the study population for any-grade events. Percentages are calculated relative to all patients in the relevant population.AEs, adverse events.aStructured AEs were derived from laboratory data. AE frequencies in the neoadjuvant systemic therapy period Arranged by AE frequency in the study population for any-grade events. Percentages are calculated relative to all patients in the relevant population. AEs, adverse events. Structured AEs were derived from laboratory data. During the NST period, 109 grade 3/4 events for 77 patients (71%) were observed in the study population versus 73 grade 3/4 events for 66 patients (65%) in the comparator population. During the complete follow-up period, 117 grade 3/4 events for 81 patients (74%) were observed in the study population versus 80 grade 3/4 events for 69 patients (68%) in the comparator population. Table 3 shows the prevalence of any type of prescription of selected supportive drugs. See Supplementary Materials (available at https://doi.org/10.1016/j.esmorw.2026.100769) for further details. Table 3Prescription prevalence of selected drugs before and after starting NSTDrug group (ATC)Comparator population (N = 101)Study population (N = 109)BaselineDuring NSTDuring and after NSTBaselineDuring NSTDuring and after NSTAnalgesics Paracetamol and anilides (N02BE)42 (42%)51 (50%)101 (100%)36 (33%)77 (71%)108 (99%) NSAIDs (M01A)29 (29%)30 (30%)100 (99%)27 (25%)49 (45%)105 (96%) Opioids (N02A)9 (9%)22 (22%)82 (81%)15 (14%)23 (21%)69 (63%) Gabapentinoids (N02BF)3 (3%)2 (2%)23 (23%)5 (5%)7 (6%)23 (21%)Dermatologics and antipruritics Dermatologic corticosteroids including antibiotic combinations (D07)11 (11%)11 (11%)47 (47%)12 (11%)21 (19%)58 (53%) Local antipruritics (D04A)0 (0%)1 (1%)1 (1%)0 (0%)0 (0%)0 (0%) Ophthalmologicals (S01)16 (16%)10 (10%)21 (21%)14 (13%)12 (11%)21 (19%)Endocrine drugs Thyroid hormones (H03AA)6 (6%)6 (6%)8 (8%)9 (8%)22 (20%)27 (25%) Systemic hydrocortisone (H02AB09)a0 (0%)0 (0%)0 (0%)0 (0%)2 (2%)6 (6%) Antidiabetics (A10)8 (8%)8 (8%)17 (17%)12 (11%)10 (9%)12 (11%) Insulins (A10A)1 (1%)3 (3%)14 (14%)2 (2%)4 (4%)6 (6%) Blood glucose lowering drugs (A10B)7 (7%)7 (7%)10 (10%)11 (10%)7 (6%)8 (7%)Gastrointestinal drugs Antiemetics and antinauseants (A04)8 (8%)101 (100%)101 (100%)15 (14%)108 (99%)109 (100%) Drugs for constipation (A06)2 (2%)14 (14%)25 (25%)7 (6%)16 (15%)28 (26%) Antipropulsives (A07D)0 (0%)8 (8%)39 (39%)0 (0%)74 (68%)80 (73%)Immunosuppressants, antiallergics, and anti-inflammatory agents Systemic glucocorticoids (H02AB)b8 (8%)101 (100%)101 (100%)11 (10%)109 (100%)109 (100%) Topical joint/muscle agents (M02)1 (1%)1 (1%)1 (1%)0 (0%)0 (0%)2 (2%) Muscle relaxants (M03)4 (4%)4 (4%)29 (29%)7 (6%)6 (6%)23 (21%) Antigout agents (M04)0 (0%)0 (0%)1 (1%)0 (0%)0 (0%)0 (0%) Bone disease agents (M05)3 (3%)2 (2%)36 (36%)9 (8%)7 (6%)34 (31%) Systemic antihistamines (R06A)19 (19%)98 (97%)100 (99%)20 (18%)107 (98%)109 (100%) Immunosuppressants (L04A)1 (1%)0 (0%)0 (0%)2 (2%)2 (2%)6 (6%)Others Anticholinesterases (N07AA)1 (1%)1 (1%)3 (3%)0 (0%)2 (2%)4 (4%) Clonidinec (C02AC01)0 (0%)5 (5%)16 (16%)0 (0%)3 (3%)4 (4%)Baseline (between 730 and 14 days before index), during NST (between 14 days before index and 14 days before operation), and during and after NST (between 14 days before index and 200 days after operation).ATC, Anatomical Therapeutic Chemical classification system; NSAID, nonsteroidal anti-inflammatory drug; NST, neoadjuvant systemic treatment.aPrimarily used to treat (suspected) adrenal insufficiency.bOften used as an antiemetic.cCan be used to treat agitation. Prescription prevalence of selected drugs before and after starting NST Baseline (between 730 and 14 days before index), during NST (between 14 days before index and 14 days before operation), and during and after NST (between 14 days before index and 200 days after operation). ATC, Anatomical Therapeutic Chemical classification system; NSAID, nonsteroidal anti-inflammatory drug; NST, neoadjuvant systemic treatment. Primarily used to treat (suspected) adrenal insufficiency. Often used as an antiemetic. Can be used to treat agitation. A total of 18 (21%) pembrolizumab-treated study patients were suspected to have experienced a grade 3/4 autoimmune disorder during NST. These included patients with hepatitis (7), nephritis (3), pneumonitis (2), colitis + adrenitis (1), colitis (1), dermatitis (1), thyroiditis (1), elevated troponin (1), and one patient with both adrenitis, thyroiditis, stomatitis, arthritis, and dermatitis. No autoimmune events were recorded during the adjuvant period or reported in the comparator population; no grade 5 AEs were identified. Pathological treatment response and TILs Seventy-four (68%) of the patients in the study population achieved pCR versus 49 (49%) in the comparator population. One patient in each group died due to rapid progression and did not undergo surgical resection of the breast tumor. Surgery specimen TILs were assessed in 19 (17%) patients in the study population and 34 (34%) patients in the comparator population; among assessed reports, 5 (26%) and 4 (12%), respectively, showed stromal TIL levels of ≥30%. Data validation Error numbers, categories, and resolution approaches are summarized in Supplementary Table S6, available at https://doi.org/10.1016/j.esmorw.2026.100769. Separate from 147 initial discrepancies attributable to manual abstraction errors (typos), 70 errors were identified in the validation study comparing algorithmically derived variables with manual EHR review. Of these, 25 were errors introduced by incorrect implementation of the intended coding logic and were identified through the validation process (including four T/N stage errors caused by challenges in algorithmically assigning stage in patients with synchronous bilateral tumors). The remaining 45 errors were related to data availability or EHR interface limitations: 12 were attributable to incomplete availability of structured pathology data during the validation period, 8 reflected true extraction discordance for weight dates and values, 2 were due to incorrect filtering during manual extraction of dermatological drugs in the EHR interface, and 23 errors were caused by outpatient prescriptions not being visible in the summarized drug field of the clinical interface. For ECOG performance status, the custom R function agreed with the initial manual abstraction for 206 of 210 patients (98%), including concordant identification of missing performance status values. The four discordant cases all represented errors in the manual abstraction. Discussion In this real-world (RW) study of neoadjuvant systemic treatment for ER-negative, HER2-normal early breast cancer, the pCR rate was higher in the study population (68%) than in the comparator population (49%), indicating improved RW effectiveness following the introduction of pembrolizumab into standard NST. However, causal interpretation of these descriptive results is limited by differences in chemotherapy exposure and potential unmeasured period-related confounding from other changes in clinical practice. We found a higher burden of toxicity after introducing pembrolizumab to SoC, with more grade 3/4 events experienced by more patients (74% versus 68% in the study and comparator populations, respectively) during or immediately after the neoadjuvant treatment. The distributions of most unstructured AEs were broadly comparable across the two populations; however, suspected severe irAEs represented a distinct toxicity signal in the study population. Among the 84 pembrolizumab-treated patients in the study population, 18 (21%) experienced a suspected severe irAE compared with 13% in the intervention arm in KEYNOTE-522. For severe AEs of any type, the distribution in the study population (74% during the complete follow-up period) was comparable to that reported in KEYNOTE-522, in which 77% of the intervention arm experienced a severe AE.4 Of note, our study population was older (median age 56 versus 49) and included more postmenopausal women (55% versus 44%) with fewer stage III and more stage I BC, but similar ECOG performance status scores, compared with the intervention arm in KEYNOTE-522. A strength of our study is the use of routinely collected EHR data from the same database that underpins clinical decision-making, extracted from the EHR system that covers all public hospital contacts for approximately half of the Danish population. Structured data had low missingness and detailed records of prescriptions, administrations, and laboratory results, enabling transparent and reproducible data flows and allowing us to report comprehensive treatment exposure estimation on an individual drug basis. Averaging treatment intensities across multiple drugs in a regimen is common and possibly reflects the combined effect of sequential and parallel drug exposures, but the transparency of summarizing individual drugs might be warranted, especially in case of toxicity reporting. By triangulating three complementary AE assessment approaches, we aimed to reduce source-specific reporting bias and provide a more comprehensive representation of patients’ experience. Due to our approval with a waiver of individual consent, we present evidence on unselected RW cohorts from participating departments, including a substantial number of patients. The comparator population provides a clear reference for temporal interpretation and future causal analyses. The validation highlights several considerations beyond identifying processing errors when using structured EHR data. Variables should be sought across relevant EHR structures, as illustrated by inconsistent weight capture in oncology-specific versus general forms. Manual review of the clinical interface is also not always an absolute reference standard, as structured EHR extracts may include data from dynamically updated fields not consistently visible during review. Finally, 147 of 217 initial discrepancies reflected manual transcription rather than extraction errors, underscoring that human review is itself error-prone. For future EHR-based RW studies, these findings highlight the need for validation across source tables, extraction logic, derivation code, and manual reference standards while distinguishing variables suited for structured extraction from those requiring curated manual review. Several limitations merit consideration. Cohort identification in EHRs is inherently challenging: despite reliance on structured pathology and treatment data, misclassification or coding errors may exclude eligible patients and threaten internal validity. The reliability of unstructured data varies: ECOG performance status is usually well documented, whereas subjective AEs are incompletely and selectively recorded, with severe events more consistently captured than mild events, leading to potentially biased toxicity estimates. Prescription patterns should not directly equate to treatment toxicity. As seen, autoimmune disorders were only reported in the study population. This may reflect true toxicity introduced by pembrolizumab, but comparisons between periods are also vulnerable to differential ascertainment and reporting bias, as clinicians were likely more attentive to immune-related symptoms after pembrolizumab implementation. This also applies to troponin, where markedly different test availability limits between-period comparison. This bias may be even more pronounced in the adjuvant period, in which less frequent oncology contact may reduce the capture of immune-related as well as other symptoms. The value gained from extracting unstructured AE data from EHR notes should be considered against the time consumed for this task and weighed against the inevitable risk of misclassification and bias. Importantly, Supplementary Table S2, available at https://doi.org/10.1016/j.esmorw.2026.100769, shows that pembrolizumab is not the only treatment exposure difference in the two populations: the study population received more carboplatin and paclitaxel as well as more pembrolizumab. This point reflects a previous finding,9 in which low RDI, driven mainly by dose reductions or discontinuations of paclitaxel and carboplatin, was correlated with a lower chance of achieving pCR after NST in triple-negative breast cancer (TNBC). Other unmeasured changes in clinical practice between the comparator and study periods may also have influenced treatment delivery, toxicity detection, and pCR. Thus, differences in effectiveness or toxicity should therefore be interpreted as period-based RW comparisons rather than causal estimates of pembrolizumab alone. Two of the six patients in the study population who were prescribed systemic hydrocortisone had confirmed immune-mediated adrenitis (2.4% of pembrolizumab-treated patients). This is comparable to KEYNOTE-522, in which 1.3% in the intervention arm experienced adrenal insufficiency. A recent RW study on early TNBC10 did, however, find 6% adrenal insufficiency in a cohort of patients treated with pembrolizumab, suggesting possible differences in populations, detection methods, or registration practices. When adherence to recommended treatment protocols cannot be assumed, metrics like relative total DI11 or RDI9 may obscure differences in drug exposure. For example, omission of a planned fourth dose may reduce both RCD and RTT to 0.75, yielding an apparent RDI of 100% despite only 75% of the planned dose being administered. This consideration is particularly relevant in our setting, in which treatment discontinuations were observed in 13%-26% of patients receiving EC and 14%-50% of those receiving carboplatin–paclitaxel. We therefore report RDI as RCD/RTT, with RCD based on administered cumulative dose and RTT corrected for the actual number of administrations received. Conclusion Patients treated with NST for early ER-negative, HER2-normal breast cancer after introduction of pembrolizumab to SoC demonstrated a higher pCR rate alongside a higher burden of toxicity, including suspected immune-mediated AEs, relative to patients treated before the introduction of pembrolizumab. Because chemotherapy delivery also differed between periods, the higher pCR rate should be interpreted as reflecting the implemented treatment strategy rather than pembrolizumab alone. This study shows that structured EHR and pathology data from hospitals treating breast cancer in eastern Denmark can support reproducible identification of patients with clearly defined clinical or pathological subtypes and detailed documentation of treatment delivery, supportive medications, laboratory-test–based AEs, and pathology-coded outcomes such as pCR. Keyword-based extraction can support simple character-string variables such as performance status and potentially menopausal status, whereas unstructured AE ascertainment remains labor-intensive and requires clinical curation. Acknowledgments Informed consent The approval included a waiver of patient consent. Ethical approval The study was approved by the Research Ethics Committee Secretary in the Capital Region (approval number R-24028034). The application was first submitted on 24 April 2024 and amended on 13 and 24 June 2024. The study was filed with the Danish Data Protection Agency (approval number P-2023-14224). Declaration of AI and AI-Assisted technologies in the writing process During the preparation of this work, the authors used GPT-5.4 and earlier OpenAI models to spellcheck and reformat already written paragraphs. After using this AI tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the article publication. Funding This work was supported by the European Union’s Horizon Europe Research and Innovation Programme [grant number 101095245] as part of the ONCOVALUE12 consortium project. Disclosure AFH has received travel, accommodations, and expense support from Daiichi Sankyo, Roche, and AstraZeneca. MAA has received research funding from Novo Nordisk and travel, accommodations, and expense support from Merck Sharp & Dohme. ASK reports personal fees from Novartis, Roche, Merck, AstraZeneca, Daiichi Sankyo, and Eli Lilly. ASK has been a national or site investigator for Daiichi Sankyo, Pfizer, Eli Lilly, and AstraZeneca. UL has received institutional research funding from Bristol Myers Squibb, Genmab, GSK, Incyte, Johnson & Johnson, Lilly, Novartis, Pfizer, and Roche and has received personal speaker’s fees from AstraZeneca. AB has received travel, accommodations, and expense support from AstraZeneca. ALU, TSP, and TAG have declared no conflicts of interest. Supplementary data Supplementary Material Supplementary Tables S1-S6 References Associated Data Supplementary Materials Supplementary Material Supplementary Tables S1-S6