Predictive Modeling of Terminal Extubation

For this secondary analysis a subset of available parameters (features) was selected for the machine learning model based on the literature study and consensus of 3 experts (PW, MS, FD). We also, a priori, decided to exclude parameters with more than 20% missing values (Fig. 1).
The following data were included in the data set:
The dependent variable was TE (yes vs. no). Independent variables used for modelling (N = 28) included:
Patients’ characteristics at baseline: study centre (binarized at 50% of terminal extubation frequency, group1 <  = 50%, group2 > 50%), age, sex, body mass index (BMI), acute physiology and chronic health evaluation II score at admission (APACHE)^{8 (link)}, chronic pre-existing medical condition (PreCond, yes/no), cardiac arrest with resuscitation before study inclusion (CPR, yes/no), admission diagnosis (3 most common categories one-hot encoded into dummy features: neurologic disorder (ADM_neuro), respiratory failure (ADM_resp), sepsis (ADM_sepsis); yes/no).
Patients’ characteristics at WLST: Glasgow coma scale (GCS), pupillary reflex (PUP, present vs. absent), cough (present vs. absent), mechanical ventilation mode (MV mode, controlled vs. supported), respiratory rate (RR, bpm), inspiratory fraction of oxygen (FiO₂%), positive end expiratory pressure (PEEP, cmH₂O), peak inspiratory pressure (PIP, cmH₂O), intubation route (route, orotracheal intubation vs. tracheostomy), mean arterial pressure (MAP, mmHg), hear rate (HR, bmp), lactate (mmol L⁻¹), arterial pH, arterial partial pressure of oxygen (PaO₂, mmHg), arterial partial pressure of carbon dioxide (PaCO₂, mmHg), total ranked circulatory drugs dose (circ_total_ranked), total ranked sedation drugs dose (sedation_total_ranked) and total ranked analgesics drugs dose at WLST (analgetics_total_ranked). Attempted donation after circulatory death was collected (DCD, yes/no).