In low-resource settings, detection of healthcare-acquired outbreaks in neonatal units relies on astute clinical staff to observe unusual morbidity or mortality from sepsis as microbiological diagnostics are often absent. We aimed to generate reliable (and automated) early warnings for potential clusters of neonatal late onset sepsis using retrospective data that could signal the start of an outbreak in an NCU in Port au Prince, Haiti, using routinely collected data on neonatal admissions. We constructed smoothed time series for late onset sepsis cases, late onset sepsis rates, neonatal care unit (NCU) mortality, maternal admissions, neonatal admissions and neonatal antibiotic consumption. An outbreak was defined as a statistical increase in any of these time series indicators. We created three outbreak alarm classes: 1) thresholds: weeks in which the late onset sepsis cases exceeded four, the late onset sepsis rates exceeded 10% of total NCU admissions and the NCU mortality exceeded 15%; 2) differential: late onset sepsis rates and NCU mortality were double the previous week; and 3) aberration: using the improved Farrington model for late onset sepsis rates and NCU mortality. We validated pairs of alarms by calculating the sensitivity and specificity of the weeks in which each alarm was launched and comparing each alarm to the weeks in which a single GNB positive blood culture was reported from a neonate. The threshold and aberration alarms were the strongest predictors for current and future NCU mortality and current LOS rates (p<0.0002). The aberration alarms were also those with the highest sensitivity, specificity, negative predictive value, and positive predictive value. Without microbiological diagnostics in NCUs in low-resource settings, applying these simple algorithms to routinely collected data show great potential to facilitate early warning for possible healthcare-acquired outbreaks of LOS in neonates. The methods used in this study require validation across other low-resource settings.