Figure 2
Algorithm pipeline for model development. Cylinders represent datasets. Boxes are algorithms or groups of algorithms. Solid arrows indicate the data/computation flow. Dashed arrows show the internal organization of groups of algorithms. Data collected from Humanitas Research Hospital (HRH) and San Raffaele Hospital (SRH) are processed by two separate groups of algorithms and merged into a single dataset. These two groups of algorithms performed data anonymization, preparation, and preprocessing. Once the datasets were merged, patients without COVID-19, jointly from the data of a previous study30, were selected to train, validate, and test the language model. Likewise, patients with COVID-19 were used for the predictive models. The preparation group algorithms select the clinically relevant predictors (selection), clean the data (cleaning) and transform the data so that for each patient, all the data are in a single row (pivoting). The preprocessing group algorithms identify the patient with COVID-19, those that died, and those that were transferred to ICU (labeling). Next, the sequence of values is converted into a single value by multiple values policies, and the out-of-range policies handle values that exceed the expected range of values.
