2.1 Study design and data sources

This report follows the Strengthening the Reporting of Observational Studies in Epidemiology Statement guidelines.^{16, 17} Two sources were used for patient cohorts in this retrospective study. First, we queried the University of Pennsylvania Health System (UPHS) Electronic Medical Records database which provides longitudinal data spanning both inpatient and outpatient settings at five hospitals including two academic medical centers in Philadelphia, a rural hospital in Lancaster, PA, and two community hospitals in urban/suburban Philadelphia.

The second source was Flatiron Health, a nationwide electronic health record (EHR)-derived de-identified database. The Flatiron Health database is a longitudinal database, comprising de-identified patient-level, structured and unstructured data, curated via technology-enabled abstraction.^{18, 19} During the study period, the de-identified data originated from ~280 cancer clinics (~800 sites of care).

Institutional Review Board approval of the study protocol was obtained before study conduct, and included a waiver of informed consent.

2.2 Eligibility

For both data sources, all patients who received 7&3 (7 days of cytarabine and 3 days of an anthracycline, either daunorubicin or idarubicin) or ven/HMA (either azacitidine or decitabine) as initial therapy (primary exposure) and had a diagnosis of acute myeloid leukemia (ICD-9 codes 205.0, 205.2, 205.3, 205.8, 205.9 or ICD-10 codes C92.0, C92.3, C92.5, C92.Z) with documented 20% blasts in bone marrow or peripheral blood were included. Patients receiving midostaurin and 7&3 were included. UPHS charts were manually reviewed to confirm that patients received 7&3 or ven/HMA as initial therapy. Any patients with acute promyelocytic leukemia, mixed phenotype or ambiguous lineage acute leukemia, or did not receive 7&3 or ven/HMA as initial therapy were excluded. Patients were excluded from analysis if they had a documented diagnosis of congestive heart failure, a recorded ejection fraction of <45% or an estimated glomerular filtration rate of <31 mL/min/1.73 m² or ECOG performance status >2 or age <60 or >75 years old. The time period for inclusion was a diagnosis made from January 2017 to January 2021 with data cut-off on June 30, 2022.

2.3 Outcomes

The primary outcome was overall survival, defined as time from diagnosis to death or censored at the end of study period on June 30, 2022. Safety outcomes were early (30- and 60-day) mortality as well as documented infections. UPHS data allowed chart review for all patients to examine febrile neutropenia, culture positive infections, laboratory abnormalities and total inpatient and outpatient hospital days before cycle 2 of therapy (“length of stay”) were exploratory outcomes. Additionally total number of “home days”; that is days without any scheduled lab visits, infusions, clinic visits, or inpatient hospitalizations were assessed. Anticipated potential effect modifiers or confounders were collected as covariates including, at time of diagnosis: demographic factors (age, race, ethnicity, practice setting), clinical factors (Hematopoietic Comorbidity Index [HCT-CI], see Tables S2c,d,²⁰ including albumin and lactate dehydrogenase [LDH]²¹ and other comorbidities including history of myelodysplastic syndrome and prior HMA treatment), disease severity (baseline blood counts, bone marrow blast counts, European LeukemiaNet—ELN—2022 risk groups²² see Table S2a, myelodysplastic syndrome related cytogenetics see Table S2b, and FLT3, IDH1, IDH2, TP53, ASXL1, RUNX1, TET2, and NPM1 mutational status).

2.4 Power

Before initial data collection, power calculation showed that 693 cases, allocated 2:1 between ven/HMA and 7&3 would have 80% power to detect a hazard ratio of 0.70 for improved OS from diagnosis to death or censoring at end of study period for 7&3, assuming an event rate of 40% at 1 year^{2, 23} with the use of a log-rank test at a two-sided significance level of 0.05 and 10% administrative censoring. By post hoc power calculation with 800 cases allocated 3:2, the minimally detected hazard ratio was 0.76.

2.5 Statistical methods

Descriptive statistics were calculated for baseline characteristics for the two treatment groups. For continuous variables, means and standard deviations for symmetric distributions, medians and inter-quartile ranges for skewed distributions, and frequencies for categorical or ordinal variables with t-test and chi-squared test to compare continuous and categorical variables are reported. Primary outcome of OS was assessed using the Kaplan–Meier method for median OS. Log rank test was used to assess survival differences between groups. Multivariable Cox proportional hazards regression analysis was used to estimate association between covariates and OS with hazard ratios and 95% confidence intervals (95% CIs). Patients were further stratified by age and cytogenetics to assess for effect modification given prior work.^{5, 24} Proportional hazards assumption was assessed using Schoenfeld residuals. Subgroups were examined in univariate analyses. Loss to follow-up, or last visit more than 3 months before data-cut-off without documented death, was assumed to be missing at random.²⁵ All tests were 2-sided, with a significance level of .05. All statistical analyses were performed in Stata.²⁶

2.6 Sensitivity analyses

This study used several sensitivity analyses to address common pitfalls of cohort studies. Missing data can severely limit multivariable Cox proportional hazards regression analysis which relies on complete cases, thus introducing bias if observations are not missing at random and limiting power. We, therefore, also conducted multiple imputation (MI), filling in missing values in multiple variables iteratively by using chained equations, a sequence of univariate imputation methods with fully conditional specification of prediction equations. Thirty imputations were used, with regression for continuous lab values, logistic regression for binary variables and ordinal logistic regression for ordinal variables.^{27, 28} Furthermore, to address confounding by indication, baseline differences between treatment groups were adjusted for, after multiple imputation, with inverse probability treatment weighting (IPTW), a propensity score-based approach to estimate marginal treatment effects. The treatment effect estimates from each imputed dataset were then combined to obtain an overall estimate.^{29, 30} In an additional sensitivity analysis a preference score, a standardizing transformation of the propensity score (logit of the propensity score – logit of Treatment A prevalence divided by Treatment B prevalence), was calculated and only patients with “equipoise” or 30%–70% chance of receiving either treatment were included as previously described.³¹ That is all patients who had less than a 30% chance of receiving either treatment based on their baseline characteristics at diagnosis were excluded in this sensitivity analysis. Additionally, to further assess the potential impact of unmeasured confounding E-values³² were calculated for the primary causal question of initial treatment after controlling for other baseline factors with MI and IPTW. E-values show the minimum strength of association on the risk-ratio scale that an unmeasured confounder would need to have with both the treatment assignment and the outcome to fully explain away a specific treatment-outcome association, conditional on the measured covariables.

3.1 Patient population

A total of 488 patients received ven/HMA (94 from UPHS, 394 from Flatiron Health) and 312 received 7&3 (85 from UPHS, 227 from Flatiron Health; Figures 1 and S1a). Median follow-up duration was 13.0 months at end of study period and median time to administrative censoring was 23.5 months. Loss to follow-up with >90 days from last visit, lab, or treatment before data-cut-off without documented death, was 5.6% overall (and similar in both groups at 7.1% for 7&3 and 4.7% for ven/HMA p = .208). Median time from diagnosis to treatment was 4.5 days for 7&3 group and 13 days for ven/HMA. There were a total of 444 death events (280 in ven/HMA and 164 in 7&3 groups). Febrile neutropenia, transfusion support, hospital length of stay and total clinic visits and infusion days were evaluable only for UPHS patients.

Patient baseline characteristics (Table 1, Figure S1) differed greatly between patients selected to receive ven/HMA versus 7&3. Patients receiving ven/HMA were older (median age 71 vs. 67 years), more likely to be treated in the community rather than an academic center (64% vs. 46% treated in the community setting), and they had more comorbidities by HCT-CI score (20% vs. 14% with a score of ≥3, see Table S2c), lower blood counts at time of diagnosis and higher rates of secondary AML (64% vs. 45%). They were also more likely to have adverse risk disease by ELN 2022 criteria (52% vs. 29%), fewer NPM1 mutations (7% vs. 25%), fewer FLT-3 ITD or FLT-3 TKD mutations (10% vs. 26%) and more high-risk mutations (including 20% vs. 4% with TP53 mutations).

Missing values in the UPHS dataset were <5% for all baseline covariates except for performance status; Flatiron Health dataset missing values ranged from 0% to 36% for covariates (e.g., sex and baseline LDH, respectively).

3.2 Overall survival

Median OS (mOS) was 14 months for all patients, 22 months for 7&3 versus 10 months for ven/HMA (HR 0.53, 95% CI 0.40–0.60, p < .0005, Figure 2A). There were 144 deaths in the 7&3 group (164/312, 39%) and 236 deaths in the ven/HMA group (280/488, 61%) by the end of study period. One year OS was 70% (95% CI 0.64–0.75) versus 44% (95% CI 0.39–0.49), and 2-year OS was 48% (95% CI 0.41–0.54) versus 25% (95% CI 0.20–0.30) for 7&3 and ven/HMA, respectively.

There was no clear subset with improved OS with either 7&3 or ven/HMA (Figures 2B and S8–S10) by univariate analysis. Patients selected to receive 7&3 had statistically significant superior overall survival in ELN adverse risk groups but similar overall survival in favorable and intermediate risk groups (Figure S7). Similar to data presented from the phase 3 VIALE-A study of venetoclax and azacitidine, there was no significant difference between favorable and intermediate risk groups by ELN 2022 criteria (Figure S7). However, there were small number of patients with core binding factor AML (11 and 8 for ven/HMA and 7&3 groups, respectively), it may not be possible to apply these findings to general populations from which ELN criteria were derived. Among patients with FLT3 mutations (n = 49 in ven/HMA group and n = 80 in 7&3 group), 81% of 7&3 patients received midostaurin, excluding these patients did not significantly change the hazard ratio between patients selected to receive 7&3 versus ven/HMA (HR 0.49–0.52).

In multivariable analysis controlling for all covariates with univariate p-value <.20 (sex, insurance, ECOG performance status, ELN risk category, type, HCT-CI score at diagnosis, mutations in ASXL1, DNMT3A, IDH, NF1, NPM1, RUNX1, TET2, TP53 and baseline white blood cell count, hemoglobin, platelets), induction choice did not affect overall survival when limited to complete cases (n = 140, HR 0.76, p-value 0.41, Table S1b).

3.3 Effect of allogeneic hematopoietic cell transplant on OS

More patients receiving 7&3 as initial therapy went on to receive transplant (31% vs. 15%, p ≤ .005). Among patients who underwent transplant, mOS was 42.3 months (Figure S2). Baseline imbalances were reflected in the two groups of patients who went on to receive transplant (e.g., 33% of ven/HMA patients who underwent transplant had TP53 mutations versus 0% of 7&3 patient who underwent transplant). Transplant status as a time-varying covariate showed that transplant improves OS as expected (HR 0.44, p < .005), and 7&3 showed higher OS compared with ven/HMA when controlling for transplant status (HR 0.52 for 7&3 vs. ven/HMA p < .005).

3.4 Sensitivity Analyses

To address missingness, after multiple imputation (MI) allowed use of all 800 patients, OS remained improved with 7&3 in multivariate cox regression analysis controlling for all covariates with univariate p-value <.20 (n = 800, HR 0.66, p-value <.001, Table S1a). To address imbalances in measured baseline covariates while accounting for missingness, we undertook a separate analysis using MI and inverse probability of treatment weighting (IPTW). After MI and IPTW, baseline covariates were well balanced with <10% absolute standardized differences between groups (Figure S3). The HR for OS, while about half as large as the unadjusted analysis, remained improved with 7&3 after balancing covariates (HR 0.71, 95% CI 0.53–0.94, p-value .026, Figure 2C). An E-value was calculated at 1.85 (E-values for upper and lower bounds of the 95% CI of hazard ratio found in MI and IPTW analysis were 1.26–2.47). An unmeasured confounder, or group of confounders, beyond all of the measured baseline co-variates balanced by MI and IPTW, with a HR of 1.85 associated with ven/HMA and death could explain away the observed HR of 0.71, but weaker confounding could not.

Given lack of overlap of probability density functions for receiving 7&3 and ven/HMA (Figure S4) there may have also been unmeasured confounding. An initially unplanned sensitivity analysis was employed to ensure the assumption of positivity was met, only patients with “equipoise” or 30%–70% chance of receiving either treatment by preference score were included as previously described.³¹ In this subgroup analysis limited to 120 out of the 488 ven/HMA (25%) and only 135 out of 312 (43%) of the 7&3 patients OS was similar at 17.5 months (HR = 1.10, 95% CI 0.75–1.6, Figure 2D).

Event free survival (EFS) was also assessed (Figure S5 and Table S2e). EFS was defined as time from the date of diagnosis to the date of induction failure, or relapse from CR, or CRi, or death from any cause; patients not known to have any of these events are censored on the date they were last examined.³³ EFS was greater in patients selected for 7&3 (HR 0.67, 95% CI 0.57–0.81, p-value <.0005), after adjusting for missingness and imbalance in baseline covariates with MI and IPTW this difference was not statistically significant (HR 0.93, 95% CI 0.75–1.17, p-value .56).

3.5 Adverse events and healthcare utilization

Thirty-day all-cause mortality was similar at 3% and 5% for 7&3 and ven/HMA, respectively (p = .20); 60-day mortality was higher for ven/HMA (6% vs. 15% p < .001 for 7&3 and ven/HMA, respectively). However, documented diagnosed infections were higher in 7&3 patients and, within the UPHS cohort where full charts could be reviewed, rates of febrile neutropenia, median transfusions during induction, and grade 3–4 electrolyte abnormalities were also higher (Table 2). Length of stay, including any admission before the next cycle of therapy, was more than twice as long for 7&3 in the UPHS cohort (31.5 vs. 15.5 days, p-value <.001). For the UPHS cohort this time reflects both readmission, prolonged admission, as well as the practice of administering 7&3 inpatient for the UPHS cohort and remaining inpatient until count recovery.

However, we also examined healthcare utilization over the entire course of therapy. “Home days” were days on which patients did not have a lab draw, infusion, office visit and were not hospitalized. The median percentage of home days was 68% for 7&3 compared with 54% for ven/HMA (Figure S6). Similarly, the total percent of days alive spent in the hospital was similar between the two groups at 10% and 12% over the entire course of treatment.