Boosting Multi-task Learning Through Combination of Task Labels - with Applications in ECG Phenotyping

Multi-task learning has increased in importance due to its superior performance by multiple different tasks simultaneously and ability perform several using a single model. In medical phenotyping, task labels are costly acquire might contain certain degree of label noise. This decreases the efficiency additional human as auxiliary when applying multi-task phenotyping. this work, we proposed an effective framework, CO-TASK, boost generating through COmbination TASK Labels. The CO-TASK framework generates without labeling effort, is robust noise, can be applied parallel with various techniques. We evaluated our CIFAR-MTL dataset demonstrated effectiveness phenotyping two large-scale ECG datasets, 18 diseases multi-label ECG-P18 echocardiogram diagnostic from electrocardiogram ECG-EchoLVH. On dataset, doubled average per-task gain model 4.38% 9.78%. With task-aware imbalance data sampler, effectively deal ratios for datasets. combined noisy annotations minor sensitivity 7.1% compared single-task while maintaining same specificity doctor on ECG-EchoLVH dataset.