Thesis Defence - Behnam Alimohammadisagvand

Tuesday, April 17, 2018, 1:00 p.m. to 3:00 p.m.

Use of Predictive Analytics to Identify Risk of Postoperative Complications among the Women Undergoing the Hysterectomy Surgery for Endometrial Cancer

Behnam Alimohammadisagvand
MSc Candidate in Health Systems
Telfer School of Management


The fourth most frequent cancer among women in Canada is endometrial cancer (EC) with about 6500 new cases diagnosed in 2015. The standard of care for the treatment of EC is a hysterectomy surgery that involves the surgical removal of the uterus. There are two types of hysterectomy surgery: open or laparotomy and minimally invasive surgery (MIS). Regardless of a type, the EC surgery may entail postoperative complications. These complications may vary from simple ones such as fever to severe complications that may result in postoperative morbidity or mortality. Identification of patients who are at risk of developing postoperative complications is a clinically important problem. It allows managing these patients differently before the surgery to prevent postoperative complications.  Hence, the primary goal of this research is to develop a data-driven predictive model for postoperative complications to be used by a surgeon at the time of consult, prior to the hysterectomy surgery. The developed model aims to predict patients at minimal or elevated risk of developing postoperative complications. For this study, we employed data of EC patients who had hysterectomy surgery at the Cancer Center of The Ottawa Hospital (TOH). The dataset including 81 attributes associated with 644 patients who underwent a hysterectomy surgery between January 1, 2012, and March 31, 2015, were analyzed in this study.  Data were collected in 3 different periods of time: 1) within 4-5 weeks prior to surgery, 2) within 24 hours after surgery, and 3) within 30 days following discharge. Ten predictive models were developed to be used prior to surgery using a set of 40 attributes. The performance of the models was assessed by F-measure, G-mean, and sensitivity as an alternative measure through a 10-fold cross-validation. A model employing PART technique with undersampling gave the best overall performance (F-measure=21.28, G-mean=55.4, and Sensitivity= 60.9) among other models prior to surgery. However, to improve the performance of the developed models, we expanded the study objectives and also developed postoperative prior to discharge. Ten predictive models were developed to be used prior to discharge using a set of 40 attributes. The PART technique with undersampling gave the best overall performance (F-measure=21.8, G-mean=56.5, and Sensitivity= 53.6) among other models prior to discharge.  As a result, it was not found a statistically significant difference between the best-performing predictive model prior to surgery and prior to discharge with regards to performance measures. Hence, shifting the time of prediction from prior to surgery to prior to discharge did not improve the performance of the predictive models. Accordingly, the PART technique with undersampling prior to surgery was selected as the best-performing predictive model to predict which patients are at a given risk category (elevated or minimal risk) of developing postoperative complications following the hysterectomy surgery.


Tuesday, April 17, 2018
1:00 p.m. to 3:00 p.m.
Telfer School of Management
Desmarais Building
DMS 4130
55 Laurier Avenue East
Ottawa, Ontario K1N 6N5
View on Google Maps

Important notice regarding photography