Understanding Birthing Mode Decision Making Using Artificial Neural Networks
| Autor: | Eugene Somoza, Martin MacDowell, Kim Brady, Albert Bocklet, Richard Fry, Kenneth Rothe |
|---|---|
| Rok vydání: | 2001 |
| Předmět: |
Decision Making
Sensitivity and Specificity Tertiary care 03 medical and health sciences 0302 clinical medicine Mode (computer interface) Pregnancy Humans Medicine 030212 general & internal medicine Artificial neural network Receiver operating characteristic analysis business.industry Singleton 030503 health policy & services Health Policy medicine.disease Obstetrics ROC Curve Predictive power Female Neural Networks Computer Artificial intelligence Medical emergency 0305 other medical science business |
| Zdroj: | Medical Decision Making. 21:433-443 |
| ISSN: | 1552-681X 0272-989X |
| DOI: | 10.1177/0272989x0102100601 |
| Popis: | Background. This study examined obstetricians’ decisions to perform or not to perform cesarean sections. The aim was to determine whether an artificial neural network could be constructed to accurately and reliably predict the birthing mode decisions of expert clinicians and to elucidate which factors were most important in deciding the birth mode. Methods. Mothers with singleton, live births who were privately insured, nonclinic, non-Medicaid patients at a major tertiary care private hospital were included in the study (N =1508). These mothers were patients of 2 physician groups: a 7-obstetrician multispecialty group practice and a physician group of 79 independently practicing obstetricians affiliated with the same hospital. A feedforward, multilayer artificial neural network (ANN) was developed and trained. It was then tested and optimized until the most parsimonious network was identified that retained a similar level of predictive power and classification accuracy. The performance of this network was further optimized using the methods of receiver operating characteristic (ROC) analysis and information theory to find the cutoff that maximized the information gain. The performance of the final ANN at this cutoff was measured using sensitivity, specificity, classification accuracy, area under the ROC curve, and maximum information gain. Results. The final neural network had excellent predictive accuracy for the birthing mode (classification accuracy = 83.5%; area under the ROC curve = 0.924; maximum information = 40.4% of a perfect diagnostic test). Conclusion. This study demonstrated that a properly optimized ANN is able to accurately predict the birthing mode decisions of expert clinicians. In addition to previously identified clinical factors (cephalopelvic disproportion, maternal medical condition necessitating a cesarean section, arrest of labor, malpresentation of the baby, fetal distress, and failed induction), nonclinical factors such as the mothers’ views on birthing mode were also found to be important in determining the birthing mode. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|