Combining an ultrasound image classifier with another classifier trained on simple clinical data increases the AUROC from 0.91 to 0.95, without the need for costly tests.
Researchers are exploring how to make automated point-of-care lung ultrasound TB screening more reliable, a tool designed for African health systems, where the disease is heavy and resources are limited. Their starting point, a ResNet network trained on images alone, already reaches an area under the ROC curve of 0.91.
The simplest fusion wins
The team compared three ways of integrating routinely collected demographic and clinical data. The result: averaging the scores of two separately trained classifiers, one on images and the other on clinical data, consistently equals or exceeds an earlier and more complex fusion. This combination achieves an AUROC of 0.95, with a sensitivity of 0.93 for a specificity of 0.76, a gain of 4 points compared to images alone.
A progressive selection of variables also makes it possible to reduce the number of questions to be asked without loss of performance. Above all, data requiring point-of-care testing brings little: the information reported by the patient or obtained by a simple measurement is sufficient for most of the time, which reduces the cost of screening.
Read the full article on the original source
Go to the source →