4 Presentations

• 
  03:30 PM - 03:55 PM

  Optimizing Screening Policies for Preventing Hospital-Acquired Infections: An MDP Model Applied to MRSA Surveillance Among Exposed Roommates in Canadian Hospitals

  • Esma Akgun, presenter, University of Waterloo
  • F. Safa Erenay, University of Waterloo
  • Sibel Alumur, University of Waterloo

  The optimal screening time and method for the exposed roommates of Methicillin-Resistant Staphylococcus Aureus (MRSA) carriers are unknown. We develop a Markov Decision Process model to determine the optimal screening method and time for individuals who have been exposed to MRSA carriers. Our model accounts for the stochastic propagation and progression of MRSA and considers quality-adjusted life years (QALYs) as the primary performance metric, while also evaluating secondary metrics such as total cost, number of colonized patients, and missed MRSA cases. Considering the spread of the infection within the room structure of a hospital makes the proposed model different from existing models, and applicable to other hospital-acquired infections. Our analyses can provide valuable insights to clinicians in developing screening policies for exposed roommates of MRSA carriers.
• 
  03:55 PM - 04:20 PM

  Effective Hospital Readmission Prediction Models using Machine-Learned Features

  • Sacha Davis, University of Alberta
  • Jin Zhang, University of Alberta
  • ILBIN LEE, presenter, University of Alberta
  • Mostafa Rezaei, ESCP Business School
  • Russell Greiner, University of Alberta
  • Finlay A. McAlister, University of Alberta
  • Raj Padwal, University of Alberta

  Hospital readmissions are one of the costliest challenges facing healthcare systems. Although 30-day readmission has been considered an important operational measure for hospitals and has gained much attention in the research community, most existing prediction models fail to predict readmissions well. Many of them use manually-engineered features, which are labor-intensive and dataset-specific. We build prediction models with features that were automatically learned by using extensive data from more than 468k patients over seven years. The data contain detailed records of hospitalizations, outpatient visits, prescriptions, physician office visits, and lab results. We use deep learning and natural language processing techniques to construct features from the complex longitudinal data. Our findings show that models using only the automatically generated features perform comparably with those using manual features. Also, we show that combining the two kinds of features improves the prediction performance substantially. Our model can be used to identify high-risk patients for whom targeted interventions may potentially prevent readmissions.
• 
  04:20 PM - 04:45 PM

  Adaptive Fuzzy Time Series Analysis for Predicting Health Expenditures in Turkey

  • Cansu Dagsuyu, presenter, Adana Alparslan Türkeş Science and Technology University
  • F. Safa Erenay, University of Waterloo
  • Ali Kokangul, University of Cukurova

  Accurate planning of health expenditures and efficient distribution of resources is critical for increasing life expectancy, decreasing mortality rates, and improving the development levels of countries. Health expenditures consist of many categories (e.g., hospital, staff, residential/ambulatory care, etc.) and are very complex to manage. We predicted health expenditures using both classical time series (TSA) and adaptive fuzzy time series analyses (AFTSA) for the categories that constitute 80% of expenditure items in 1999-2021 Turkish Statistical Institute (TUIK) data. This comparison illustrates that AFTS models perform better than TSA models in terms of mean absolute, mean squared, root mean square, and mean absolute percentage error. Accurately forecasting health expenses contributes to the comparative assessment of nations, which guides the attempts for improving the financial/organizational structures of the national health system.
• 
  04:45 PM - 05:10 PM

  Application of Machine Learning for Long-term Graft Survival Prediction of Liver Transplant

  • Dionne Aleman, University of Toronto
  • Louis-Martin Rousseau, Polytechnique Montreal & CIRRELT
  • Hamed Shourabizadeh, presenter, University of Toronto
  • Mamatha Bhat, University of Toronto
  • Katina Zheng, University of Toronto

  We develop machine learning models to predict long-term graft survival in liver transplant patients using only pre-transplant variables in the Scientific Registry of Transplant Recipients (SRTR) database. We included a total of 79 numerical and categorical variables. Two different target variables were used to predict graft survival: i) 20-year survival as a binary target variable, ii) survival function as the target variable. Then, two classification and survival machine learning algorithms were trained: random forest, and XGBoost Survival Embeddings. The performance of the algorithms was evaluated using the area under the curve (AUC).
  We investigate the effectiveness of transferring ML models from one centre to another by dividing the SRTR dataset into 11 subsets based on the 11 US regions (OPTNs). The models are trained on each OPTN and tested on the other ten. The results show that the trained models do not significantly impact the performance of the algorithms, indicating that it is possible to use pre-trained models from other centres. However, most transplant centres have different sets of variables. This reduces the efficiency of using a model trained on another database. Therefore, using local single-centre datasets ensure that the models are trained on the exact same dataset without having to remove or adjust the variables.