2030. How Machine-learning Technique Using Artificial Neural Network Determines Whether The Fever Is Actually Related To The Bacteremia.
Session: Poster Abstract Session: Diagnostics: Biomarkers and Novel Approaches
Saturday, October 6, 2018
Room: S Poster Hall
Posters
  • 2018_IDSA_AI_가로-converted.pdf (290.7 kB)
  • Background: By applying machine-learning based algorithm using artificial intelligence to massive medical data, we are trying to build a real-time monitoring system for prediction of diseases to support accurate and efficient clinical decision making in time. In the previous study, we presented a model for predicting bacteremia using Bayesian statistical approach. Now, we have developed various machine-learning technique based prediction model to achieve better prediction performance.

    Methods: We retrospectively analyzed 13,402 febrile patients who were admitted to Gangnam Severance Hospital, a tertiary center in Seoul, South Korea. The training data were 11,061 patients with admission date from Jul 2008 to Aug 2011, and validation data were 2,341 patients from Sep 2011 to Feb 2012.

    The primary outcome was bacteremia, and the training data were analyzed to make prediction model with conventional Bayesian approach, Support Vector Machine (SVM), Random Forest (RF) and multi-layer perceptron (MLP), a representative artificial neural network (ANN) model, respectively. The performance of prediction was assessed based on the area under the curve (AUC) and sensitivity from validation data. We used twenty clinical variables for predictors of bacteremia same as Bayesian approach. The difference from the previous model was that each variable had been stratified, but in this study, they were trained by continuous number as it is.

    Results: A total of 1,538 bacteremia episodes were identified from 13,402 febrile patients. The AUC of bacteremia prediction performance in SVM model was lowest with the result of 0.699 (95%CI; 0.687-0.700), even though it was 0.7 in conventional Bayesian statistical method. The highest results were 0.732 (95% CI; 0.722-0.733) in RF model and in MLP with 128 nodes of hidden layer model, the AUC was 0.719 (95% CI; 0.712-0.728) and in MLP with 256 nodes, it was 0.727 (95%CI; 0.713-0.727). In comparison with sensitivity, MLP models (0.810, 95%CI 0.772-0.747 in 128 nodes, 0.810, 95% CI, 0.782-0.837 in 256 nodes) were the highest but in RF model, the sensitivity was the lowest.

    Conclusion: Compared with conventional statistical model, ANN based bacteremia prediction model-MLP showed better predictive value. In order to improve the performance of prediction, further larger amount of clinical data is needed to be analyzed.

    Kyoung Hwa Lee, MD1, Seul Gi Yoo, MD1, Da Eun Kwon, MD1, Soon Young Park, MS1, Jae June Dong, MD2, Myunghun Chae, MS3, Choongki Min, MS4, Jaewoo Kang, MS3 and Young Goo Song, MD, PhD1, (1)Division of Infectious Diseases, Department of Internal Medicine, Gangnam Severance Hospital, Yonsei University College of Medicine, Seoul, Korea, Republic of (South), (2)Department of Family Medicine, Gangnam Severance Hospital, Yonsei University College of Medicine, Seoul, Korea, Republic of (South), (3)Selvas Artificial Intelligence Incoporate, Seoul, Korea, Republic of (South), (4)Selvas Artificial Intelligence incoporate, Seoul, Korea, Republic of (South)

    Disclosures:

    K. H. Lee, None

    S. G. Yoo, None

    D. E. Kwon, None

    S. Y. Park, None

    J. J. Dong, None

    M. Chae, None

    C. Min, None

    J. Kang, None

    Y. G. Song, None

    Findings in the abstracts are embargoed until 12:01 a.m. PDT, Wednesday Oct. 3rd with the exception of research findings presented at the IDWeek press conferences.