Mo1909 Prognostic Significance of Bioelectrical Impedance Analysis Defined Body Composition in Upper Gastrointestinal Cancer
| Autor: | Wyn G. Lewi, Charlotte E. Thomas, Gary Howell, Jolene Witherspoon, Paul A. Blake, Llion Davies, Rachael Barlow, Andrew J. Beamish, Alex Karran |
|---|---|
| Rok vydání: | 2014 |
| Předmět: |
medicine.medical_specialty
education.field_of_study Boosting (machine learning) Hepatology business.industry Mortality rate Data classification Population Gastroenterology Logistic regression Internal medicine Test set Statistics medicine Stage (cooking) education business Bioelectrical impedance analysis |
| Zdroj: | Gastroenterology. 146:S-689 |
| ISSN: | 0016-5085 |
| DOI: | 10.1016/s0016-5085(14)62499-9 |
| Popis: | years diagnosed with colorectal cancer between 1985 and 2004 were included in the study. After preprocessing and joining the raw data across both registries, the population consisted of 135,000 data records, containing 60 usable variables for potential inclusion in the model. After dividing the test set into training and testing sets, we used the information gain ration methodology, as well as a novel oversampling balancing technique that generates synthetic data points to account for the loss of information resulting from death of patients between years 1 and 5 post-diagnosis. We selected 11 predictive attributes, including tumor size and extension, lymph node involvement, regional nodes and primary site involvement, stage, histologic type, and demographic factors including age, gender, and place of residence. Then, experiments were run on 25 data classification schemes consisting of basic classifiers (trees, functions, and logistic regression) and boosting meta-classifiers. Using 10-fold crossvalidation, the validity of each of these classification schemes was tested. Results: Combining basic and meta classifiers chosen from the 25 candidate schemes resulted in highly accurate prediction of survival rates using only 11 of the most predictive patient data features from 60 potential features. Namely, the model achieved 89.5% accurate prediction of the 1-year post-diagnosis survival rates of CRC patients, and 86.2% accurate prediction of the 5-year post-diagnosis CRC survival rates. Conclusion: By combining basic and meta-classifiers and using a novel combined database larger than others in preceding studies, a model was developed using just 11 of 60 potential variables to predict mortality rates of CRC patients 1 and 5 years post-diagnosis, with accuracy of approximately 90% and 86%, respectively. The insights generated by this model could aid diagnosed patients and clinicians greatly in developing cancer treatment and surveillance plans. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|