Fig. 5: Network visualization and statistics of the feature significance for the model prediction. | npj Digital Medicine

Fig. 5: Network visualization and statistics of the feature significance for the model prediction.

From: iCOVID: interpretable deep learning framework for early recovery-time prediction of COVID-19 patients

Fig. 5

a Feature map visualization of iCOVID corresponding to four representative patient examples (color masks the significant regions for the prediction, with the spectrum from blue to red associated with low-to-high significance). b Heatmap of the average significance of each feature and the recovery days, revealing that the biomarkers AM, TP, and HG are significant for the prediction of recovered patients, whereas the comorbidities SK, ARDS, and DB are more significant for the prediction of deceased patients (color indicates the significance of each feature for the prediction, with the spectrum from dark purple to yellowish-white associated with low-to-high significance). c The average significance of the top 15 clinical features, i.e., AM albumin, HG hemoglobin, EPC expectoration, TP total protein, DB diabetes, ARDS acute respiratory distress syndrome, SK shock, DH diarrhea, SN soreness, FV fever, CGH cough, LDH lactate dehydrogenase, PA poor appetite, CCBD chest congestion/breathing difficulty. d Pearson correlation analysis demonstrates that the above-mentioned features are indeed highly related to the recovery time of COVID-19 patients (p value < 0.001, except for EPC, DH, and SK).

Back to article page