We identified all patients age ≥18 years admitted to four hospitals in eastern Massachusetts between May 2017 and February 2021 who were treated with antibiotics with a stated indication of pneumonia. Study hospitals included two large tertiary referral hospitals (Brigham and Women’s Hospital, Massachusetts General Hospital) and two community hospitals (Faulkner Hospital, Newton Wellesley Hospital). Granular clinical data were extracted from an Enterprise Data Warehouse populated by the electronic health record system shared by all study hospitals (Epic Systems, Verona, WI) including demographics, diagnosis codes, locations, services, vital signs, laboratory values, medications, and antibiotic indications. Providers are required to specify an indication whenever new antibiotics are ordered for inpatients. This is done within the antibiotic order screen in the electronic health record by selecting from a structured list of options that includes pneumonia. Providers can also enter a free text indication. Validation studies suggest that the indications specified by clinicians accurately reflect their working diagnoses at the time of prescription as specified in their concurrent clinical notes.[10, 23]

We limited the population to patients with median daily oxygen saturation levels of ≥95% without supplemental oxygen on both the first and second calendar days of antibiotics. Median oxygen saturation levels were calculated using all recorded values on a calendar day. We excluded patients with positive blood cultures (other than common skin contaminants), missing vital signs on the first day of antibiotics, or discharge diagnosis codes for empyema, cystic fibrosis, or bronchiectasis (E84, J85, J86) since standard of care for these conditions includes longer antibiotic courses. If a patient was admitted more than once during the study period then a single admission was selected at random. Missing laboratory values were imputed from the closest measured value within 2 days before or after the first day of antibiotics but if no measurements were available then missing values were imputed as normal.

We compared outcomes for patients treated with 1-2 days of antibiotics versus those treated with 5-8 days of antibiotics by propensity-matching those treated for 1-2 days to those treated 5-8 days using a caliper of 0.1 times the standard deviation of the estimated logit propensity score.[24] Post-discharge antibiotics were included when calculating duration of antibiotics. We chose these two intervals (1-2 days and 5-8 days) to provide clear separation in treatment durations between groups and because we hypothesized clinical equipoise amongst clinicians on whether patients with normal oxygenation levels have pneumonia or not: we reasoned that clinicians who stopped antibiotics within 1-2 days likely decided against pneumonia whereas those who treated for 5-8 days opted in favor of pneumonia. We recognize, however, that clinicians may be more likely to stop antibiotics if evidence for pneumonia is more equivocal. We therefore propensity matched patients treated 1-2 days to patients treated 5-8 days using an extensive array of granular clinical parameters including demographics, comorbidities, vital signs, laboratory values, culture orders, respiratory viral studies, and baseline medications. We also limited the analysis to patients who survived at least 5 days counting from the first day of antibiotics in order to assure that antibiotic courses of less than 5 days were due to clinical choice, not early death, and thus mitigate survivor bias in favor of those treated with longer courses.

Propensity scores for treatment interval were calculated by stratifying patients by hospital and then within each hospital using patient’s demographics (age, race, sex), clinical service on the first day of antibiotics (medicine, neurology, obstetrics, oncology, surgery, cardiology, cardiac surgery), presence in an intensive care unit on the first day of antibiotics, comorbidities (congestive heart failure, renal failure, liver disease, metastatic cancer, solid tumor, diabetes, cerebrovascular disease, pulmonary circulation disorder), vital signs on the first day of antibiotics (maximum temperature, median respiratory rate, maximum heart rate, minimum systolic blood pressure), laboratory values on the first of antibiotics (maximum white blood cell count, minimum hematocrit, minimum platelets, maximum creatinine, minimum sodium, maximum glucose, maximum alanine aminotransferase, maximum bilirubin, maximum albumin, maximum international normalized ratio), vasopressor use on the first day of antibiotics, antibiotic exposure for other indications between admission and the first day of antibiotics for pneumonia, number of days from admission until first day of antibiotics, whether blood cultures were obtained, whether a sputum culture was obtained, whether sputum cultures were positive for potentially pathogenic organisms (i.e. organisms other than coagulase-negative Staphylococci, Enterococcus, and Candida), and whether any assays for respiratory viruses were positive (including SARS-CoV-2, influenza, respiratory syncytial virus, parainfluenza, rhinovirus, human metapneumovirus, and adenovirus). Patients’ comorbidities were derived using the methods of Charlson and Elixhauser.[25, 26] We used the median respiratory rate rather than the minimum or maximum given that this parameter is prone to wide fluctuation over the course of a day depending on patients’ activities and clinical interventions.

We compared hospital death, hospital discharge 30-day mortality, 30-day readmission, and 90-day positive Clostridioides difficile counting from the first day of antibiotics using sample proportions, and total antibiotic days, total antibiotic-free days alive, and total hospital-free days alive within 30 days of the first day of antibiotics using sample means. We calculated P-values for comparisons in the total and matched cohorts using two-sample t-tests.

We further estimated measures of association for short vs conventional treatment groups amongst propensity-matched patients using Fine and Gray regression for propensity-score matched data [27] to estimate subdistribution hazard ratios for hospital death and discharge, logistic regression to estimate odds ratios for 30-day mortality, 30-day readmission and 90-day positive Clostridioides difficile toxin assays. We used linear regression to estimate mean differences for total antibiotic days, total antibiotic-free days alive, and total hospital-free days alive within 30-days of the first day of antibiotics. We generated 95% confidence intervals for odds ratios and mean differences using cluster-robust standard errors to account for the paired nature of the data.

We performed sensitivity analyses limited to patients with community acquired pneumonia (first day of antibiotics on hospital day 1-2), hospital-acquired pneumonia (first day of antibiotics on hospital day ≥3), patients in whom sputum cultures were obtained, patients with positive sputum cultures, and patients with discharge diagnosis codes for pneumonia (J13-J18). We also did a sensitivity analysis in which we matched patients using their clinical parameters from the second day of antibiotics since decision making regarding duration of antibiotics may have been made on the basis of patients’ clinical parameters on the second day of treatment. All calculations were performed using R version 4.03. The study was approved by the Mass General Brigham Institutional Review Board.

There were 39,752 patients started on antibiotics with a stated indication of pneumonia during the study period. Of these, 10,012 had a median oxygen saturation of ≥95% without supplemental oxygen on the first and second days of antibiotics. After applying the study exclusion criteria and limiting to patients treated for 1-2 or 5-8 days there were 5,762 patients remaining (2,871 treated for 1-2 days and 2,891 treated for 5-8 days). After propensity matching, 4,478 patients were included in the primary analysis (2,239 in each group). Characteristics of patients before and after propensity matching are shown in Table 1. Within the propensity matched set, average age was 66 years, 55% were male, 68% were white, 73% were on an internal medicine service, 23% had cancer, 28% had diabetes, 26% had renal failure, and 26% had congestive heart failure. Antibiotics for pneumonia were started within the first 2 days of hospitalization in 3466/4478 (77%) and on hospital day 3 or later in in the remainder. Standardized differences between patients treated for 1-2 days versus those treated for 5-8 days were <0.1 on all patient characteristics after propensity matching.[28]

In this large propensity-matched analysis we did not detect any statistically significant differences in mortality or readmissions amongst patients treated for possible pneumonia who had oxygen saturations ≥95% without supplemental oxygen regardless of whether they were treated with 1-2 days of antibiotics versus 5-8 days of antibiotics. Our findings suggest that deliberate review of oxygenation levels may be a practical strategy to help clinicians identify patients on antibiotics for possible pneumonia who are promising candidates for early discontinuation of antibiotics.

The lack of apparent benefit of conventional antibiotic course durations for patients with possible pneumonia but preserved oxygenation likely reflects a combination of factors. Some patients may not have had pneumonia at all, a frequent phenomenon in the hospitalized population given the non-specific nature of the clinical symptoms, signs, laboratory tests, and radiographs used to diagnose pneumonia.[6-11] Other patients may have had pneumonia but with a viral rather than a bacterial etiology. One third to one half of community-acquired pneumonias in hospitalized patients and one fifth of hospital-acquired pneumonias are attributable to viruses.[29-31] A further set of patients may indeed have had bacterial pneumonia but mild cases with minimal parenchymal invasion and / or very good physiologic reserves as suggested by their preserved oxygenation levels such that they were able to recover with very brief treatment courses.

Our results are compatible with the growing literature demonstrating the safety of brief antibiotic courses for community acquired pneumonia. Two randomized trials comparing the safety of 3 versus 8 days of antibiotics for patients hospitalized with moderate to severe pneumonia reported no differences in outcomes between groups.[32, 33] A third trial reported a similar finding amongst ventilated patients with pulmonary infiltrates and low clinical pulmonary infection scores who were randomized to stopping antibiotics after 3 days versus usual care.[22] Our study builds upon these trials by demonstrating the generalizability of these findings to real-world populations without the enrollment exclusion criteria applied in these trials, the potential utility of using oxygenation levels to identify candidates for early discontinuation of antibiotics, and the potential safety of stopping antibiotics after less than three days amongst the subset of patients with suspected pneumonia but preserved oxygenation levels.

Limitations of our study include the observational design and thus the possibility of unmeasured confounding by indication. Clinicians may have been more likely to stop antibiotics early in patients who appeared to be more stable or in whom other information emerged that pointed to a non-bacterial or non-infectious process. We attempted to control for confounding by indication, however, by propensity-matching patients using an extensive array of clinically detailed variables including patients’ demographics, comorbidities, vital signs, laboratory values, bacterial and viral testing, and prior treatments. Despite the large number and granularity of the clinical parameters included in our propensity scores we were able to achieve a close match without a large loss in sample size. Indeed, it is notable that even before matching, patients were almost perfectly divided into short course and conventional course groups with small standardized differences for the majority of patient characteristics. This suggests there is some degree of clinical equipoise at the bedside as to whether patients with possible pneumonia but preserved oxygen saturations need antibiotics or not. Nonetheless, the finding that patients treated with shorter courses may have been discharged sooner may be due to residual confounding.

Other limitations of our study include our focus on four hospitals from a single region of one country. This may limit generalizability. It is also possible that treating clinicians misspecified indications for some patients leading to contamination of the study sample with patients without pneumonia. Prior validation studies, however, have suggested that the indications provided by clinicians when they prescribe antibiotics accurately reflect their clinical impressions at the time of specification.[10, 23] In addition, results were consistent in the sensitivity analyses restricted to patients with a higher probability of having true pneumonia, namely those with positive sputum cultures and those with discharge diagnosis codes for pneumonia. Post-discharge antibiotic durations were calculated from discharge prescriptions but may not reflect what patients took if patients failed to fill prescriptions, stopped their courses early, or received extensions from their providers. Finally, propensity score matching does not accommodate time-varying confounders that are affected by past treatment; for example, some residual bias in our estimates is possible if time-updated lab measures or vital signs suggesting disease progression affected clinicians’ decisions to stop or continue antibiotics and were also, themselves, affected by past antibiotic decisions.[34] Our findings were consistent on a sensitivity analysis where we used patients’ clinical parameters on the second day of treatment rather than the first but bias may still occur even when such measures are included in the propensity score model. Emerging causal inference methods can better accommodate these more complex data interactions and are a promising route for future observational studies of time-varying treatment strategies.[35]

In summary, we found no evidence of harm and possible benefit with 1-2 versus 5-8 days of antibiotics to treat possible pneumonia in patients who had oxygen saturations ≥95% on ambient air. This analysis suggests a potentially powerful strategy to help clinicians to easily identify a subset of patients with possible pneumonia in whom early discontinuation of antibiotics may be safe. The potential population that could benefit from this strategy is large. Our study and others suggest that a quarter to a third of hospitalized patients treated for possible pneumonia may fall into this category.[10, 20, 21] Early discontinuation of antibiotics started for possible pneumonia in hospitalized patients with normal oxygenation merits testing in a prospective randomized trial to confirm or refute the feasibility, safety, and utility of this strategy for reducing unnecessary antibiotic use.