Use of a gas-operated ventilator as a noninvasive bridging respiratory therapy in critically Ill COVID-19 patients in a middle-income country.

During the COVID-19 pandemic, there was a notable undersupply of respiratory support devices, especially in low- and middle-income countries. As a result, many hospitals turned to alternative respiratory therapies, including the use of gas-operated ventilators (GOV). The aim of this study was to describe the use of GOV as a noninvasive bridging respiratory therapy in critically ill COVID-19 patients and to compare clinical outcomes achieved with this device to conventional respiratory therapies. Retrospective cohort analysis of critically ill COVID-19 patients during the first local wave of the pandemic. The final analysis included 204 patients grouped according to the type of respiratory therapy received in the first 24 h, as follows: conventional oxygen therapy (COT), n = 28 (14%); GOV, n = 72 (35%); noninvasive ventilation (NIV), n = 49 (24%); invasive mechanical ventilation (IMV), n = 55 (27%). In 72, GOV served as noninvasive bridging respiratory therapy in 42 (58%) of these patients. In the other 30 patients (42%), 20 (28%) presented clinical improvement and were discharged; 10 (14%) died. In the COT and GOV groups, 68% and 39%, respectively, progressed to intubation (P ≤ 0.001). Clinical outcomes in the GOV and NIV groups were similar (no statistically significant differences). GOV was successfully used as a noninvasive bridging respiratory therapy in more than half of patients. Clinical outcomes in the GOV group were comparable to those of the NIV group. These findings support the use of GOV as an emergency, noninvasive bridging respiratory therapy in medical crises when alternative approaches to the standard of care may be justifiable.

Neural networks memorise personal information from one sample.

Deep neural networks (DNNs) have achieved high accuracy in diagnosing multiple diseases/conditions at a large scale. However, a number of concerns have been raised about safeguarding data privacy and algorithmic bias of the neural network models. We demonstrate that unique features (UFs), such as names, IDs, or other patient information can be memorised (and eventually leaked) by neural networks even when it occurs on a single training data sample within the dataset. We explain this memorisation phenomenon by showing that it is more likely to occur when UFs are an instance of a rare concept. We propose methods to identify whether a given model does or does not memorise a given (known) feature. Importantly, our method does not require access to the training data and therefore can be deployed by an external entity. We conclude that memorisation does have implications on model robustness, but it can also pose a risk to the privacy of patients who consent to the use of their data for training models.