Table 3.

Model performance during internal validation, prospective validation, and external validation.

	Accuracy	Precision	Recall	F1-score
Internal validation (N = 680)
Task 1: normal vs. abnormal	0.8368	0.8325	0.8340	0.8332
Task 2: crackles vs. wheezing	0.8367	0.8382	0.8360	0.8363
Task 3: normal vs. crackles	0.8094	0.8007	0.7938	0.7966
Task 4: normal vs. wheezing	0.9042	0.9053	0.8863	0.8936
Prospective validation (N = 90)
Task 1: normal vs. abnormal	0.8222	0.7920	0.8122	0.8000
Task 2: crackles vs. wheezing	0.6774	0.6804	0.6774	0.6761
Task 3: normal vs. crackles	0.6780	0.6972	0.6849	0.6746
Task 4: normal vs. wheezing	0.8136	0.8298	0.8191	0.8127
External validation (N = 782) (ICBHI 2017 pediatric data)
Task 1: normal vs. abnormal	0.835	0.649	0.171	0.793
Task 2: crackles vs. wheezing	0.764	0.828	0.707	0.764
Task 3: normal vs. crackles	0.911	1	0.031	0.871
Task 4: normal vs. wheezing	0.915	1	0.187	0.888

ICBHI International Conference on Biomedical and Health Informatics 2017 Challenge Respiratory Sound Database.