Decrease of haemoconcentration reliably detects hydrostatic pulmonary oedema in dyspnoeic patients in the emergency department - a machine learning approach.

Background: Haemoglobin variation (ΔHb) induced by fluid transfer through the intestitium has been proposed as a useful tool for detecting hydrostatic pulmonary oedema (HPO). However, its use in the emergency department (ED) setting still needs to be determined.

Methods: In this observational retrospective monocentric study, ED patients admitted for acute dyspnoea were enrolled. Hb values were recorded both at ED presentation (T₀) and after 4 to 8 h (T₁). ΔHb between T₁ and T₀ (ΔHb_T1-T0) was calculated as absolute and relative value. Two investigators, unaware of Hb values, defined the cause of dyspnoea as HPO and non-HPO. ΔHb_T1-T0 ability to detect HPO was evaluated. A machine learning approach was used to develop a predictive tool for HPO, by considering the ability of ΔHb as covariate, together with baseline patient characteristics.

Results: Seven-hundred-and-six dyspnoeic patients (203 HPO and 503 non-HPO) were enrolled over 19 months. Hb levels were significantly different between HPO and non-HPO patients both at T₀ and T₁ (p < 0.001). ΔHb_T1-T0 were more pronounced in HPO than non-HPO patients, both as relative (-8.2 [-11.2 to -5.6] vs. 0.6 [-2.1 to 3.3] %) and absolute (-1.0 [-1.4 to -0.8] vs. 0.1 [-0.3 to 0.4] g/dL) values (p < 0.001). A relative ΔHb_T1-T0 of -5% detected HPO with an area under the receiver operating characteristic curve (AUROC) of 0.901 [0.896-0.906]. Among the considered models, Gradient Boosting Machine showed excellent predictive ability in identifying HPO patients and was used to create a web-based application. ΔHb_T1-T0 was confirmed as the most important covariate for HPO prediction.

Conclusions: ΔHb_T1-T0 in patients admitted for acute dyspnoea reliably identifies HPO in the ED setting. The machine learning predictive tool may represent a performing and clinically handy tool for confirming HPO.