by Barbara Boughton
San Francisco, CA - Artificial neural networks—teachable software designed to mimic the human brain—could one day be used to diagnose endocarditis related to implanted cardiac rhythm devices.
In a proof-of-concept study, investigators from the Mayo Clinic in Rochester, MN devised a neural network that could correctly identify 72 of 73 generator pocket infections and 12 of 13 cardiac rhythm management device-related infective endocarditis (CRMD-IE), according to research presented here at the American Society for Microbiology 2009 Interscience Conference on Antimicrobial Agents and Chemotherapy.
If proven, the software could benefit some patients by helping them avoid the risk, discomfort, and expense of transesophageal echocardiography (TEE), according to lead author Dr M Rizwan Sohail (Mayo Clinic). "In some cases, it might enable us to correctly diagnose patients without performing TEE or to select patients who might benefit from additional testing," Sohail said. "The result would be reduced healthcare costs and improved outcomes."
The researchers used data from 189 patients admitted to the Mayo Clinic with a diagnosis of CRMD infection to "train" the software to recognize variables that signaled infection. Of the patients, 23% actually had CRMD-IE.
In testing the artificial neural network, the researchers used three configurations of clinical variables to test how accurate the software could be. They found that the most reliable variables were blood-culture results, presenting signs and symptoms, tenderness at the infection site, ejection fraction, and fever. By using these variables, they increased the overall sensitivity of the artificial neural network to 86.4%, Sohail said.
In developing their model, they focused primarily on clinical features of CRMD infection, rather than laboratory parameters, because these findings are readily available to clinicians. However, the inclusion of positive blood cultures and presenting signs and symptoms were the variables that most improved the sensitivity of their model, Sohail said.
The scientists used the final neural network on six cases with an "unknown" diagnosis. They found that the trained software correctly identified all four cases without CRMD-IE (pocket-only infections) but missed one of the two CRMD-IE cases.
The scientists are continuing to work with the artificial neural network and have already used more than 400 cases to continue "training" the software. They also plan to expose it to 200 cases with unknown diagnoses to see how reliable the software is. If the model proves reliable, the artificial neural network might eventually be used to diagnose other types of infection, such as pneumonia, Sohail said.
"In training the artificial neural network, you have to be very careful about what data you feed into it, and expose it gradually to different variables," he said. Like humans, too much extraneous data can cause the artificial neural network to "become distracted," he said.
"These results are very encouraging," commented Dr Vance G Fowler Jr (Duke University, Durham, NC). "If appropriately validated, it could help clinicians make a sometimes-difficult diagnosis."
Yet Fowler pointed out that the artificial neural network correctly identified only a handful of CRMD-IE cases, while it was more successful in pinpointing generator pocket infections, which are fairly easy for clinicians to recognize, he said. "But it's a novel and innovative approach, and it clearly warrants further investigation," he said.
Source
Sohail MR, Saadah LM, Uslan DZ, et al. Using artificial neural networks to predict cardiac rhythm management device-related infective endocarditis. American Society for Microbiology 2009 Interscience Conference on Antimicrobial Agents and Chemotherapy; September 12, 2009; San Francisco, CA. Abstract K-268.
0 comments:
Post a Comment