Data Analytics in Healthcare 2020

Consecutive Surgeries with Complications: The Impact of Scheduling Decisions.

Adam Diamant, Schulich School of Business, York University

Abstract: Few studies have investigated the quality implications of workers exercising discretion in scheduling their own work, i.e., deciding which tasks to perform and when, far in advance of when the tasks are actually completed. To this end, our study focuses on general surgeons who are given privileges at a hospital to perform elective surgery, i.e., a medical procedure that is scheduled in advance of its performance because it does not involve an acute medical emergency. For each surgical day allocated to them by the hospital, surgeons select which patients to assign and in what order the corresponding surgeries will be performed. We track both health outcomes (whether there is a complication) and the length-of-stay (LOS) of patients who undergo elective surgery and determine how these measures are affected by the order in which surgeons operate. More specifically, by analyzing a large data set of 29,169 surgeries performed by 111 surgeons from 2005 to 2015 in a major hospital network, we determine how the scheduling and sequencing of elective surgeries by surgeons impacts the rate of surgical complications and patient LOS. We find that surgeries following those that experienced a complication were more likely to experience a complication and these patients were also more likely to be admitted to the hospital for a longer duration. The increased complication risk and LOS was not affected by scheduling greater slack time between surgeries for rest, recovery, and reflection, nor was the phenomenon localized to a few problematic individuals. However, we do find support for the notion that scheduling surgeries that require different techniques does mitigate these risks. Our results illustrate how operational decision-making can affect clinical outcomes. It also has implications for the theory and practice on how best to schedule complex, knowledge intensive work and how best to provide administrative support.

Watch a video of the presentation →
Download a copy of the presentation slides →

When Doctors are Most Needed in the Emergency Room: Risk-Adaptive Physician Shift Scheduling.

Yichuan (Daniel) Ding, McGill University

Abstract: In this paper we optimize Emergency Department (ED) physicians' staffing by locally adjusting the shift schedule in order to minimize patients' total waiting cost. We first develop a structural model to infer a patient’s waiting cost perceived by a typical ED physician based on the historical patient sequencing. The estimation results show that a patient’s waiting cost is a piece-wise convex function in her cumulative waiting time. We provide two indices where we find that the essence of this problem lies in comparing the patient mix in the expected busy periods. We characterize our adjustment policy through pathwise analysis of the historical data and offer instructions of implementation in a local ED.

Watch a video of the presentation →
Download a copy of the presentation slides →

Factors Affecting Non-urgent Patient Visits to Emergency Departments: A Discrete Choice Experiment

Speaker: Shrutivandana Sharma, Singapore University of Technology and Design

Abstract: Non-urgent patient visits have been one of the major factors behind rise in emergency department (ED) visits and ED crowding. This work investigates factors that influence non-urgent patients’ choices between ED and General Practitioner (GP). A discrete choice experiment (DCE) survey was developed to elicit patients’ preferences for ED and GP. Responses from 849 respondents recruited from a public hospital in Singapore were included in the study. The survey responses were used to develop patient choice models using latent class multinomial logistic regression. In addition to quantifying the influence of general ED/GP attributes on patients’ preferences, these choice models also quantity the influence of a new GP-referral discount scheme that was introduced by a public hospital in Singapore to encourage non-urgent patients to first visit GPs instead of directly visiting the ED. Our findings suggest that waiting time, test facilities, out-of-pocket payment as well as GP-referral discount significantly influence patients’ preferences for ED and GP. In addition, patients are heterogeneous in their preferences, particularly with respect to availability of test facilities and cost of care. We map heterogeneity of patients’ preferences to patients’ demographics and their perception of the criticality of their medical condition. We find that patients with more than 40 years age, part-time employment, shorter travel time to ED, and perception of their condition as “critical enough to go to ED directly” are more sensitive to test facilities and have an inherent preference for ED. The relative importance of factors quantified by the choice models suggest various countermeasures that can be beneficial for reducing non-urgent ED visits.

Presentation unavailable.

Can customer arrival rates be modelled by sine waves?

Ningyuan Chen, University of Toronto

Abstract: Patient arrival patterns observed in the real world typically exhibit strong seasonal effects. It is therefore natural to ask: Can a nonhomogeneous Poisson process with a rate that is the simple sum of sinusoids provide an adequate description of reality? If so, when is it advantageous to use a sinusoidal rate in practice? We introduce a new sinusoidal rate model from the statistical learning literature to fit the arrivals data, and also to estimate the periodicities in the arrival pattern.

We empirically investigate the questions using patient arrivals to an emergency department. We find that the model is consistent with the arrivals data. For operational decision-making, the sinusoidal rate model is better than commonly used piecewise smooth fits when the arrival rate is not periodic. When the arrival rate is periodic, the sinusoidal model does no worse. Taken together, the flexibility and tractability of the sinusoidal specification suggest that it is a worthy workhorse model for time-varying arrival processes.

Watch a video of the presentation →
Download a copy of the presentation slides →

Keynote Speaker: An Examination of Early Transfers to the ICU Based on a Physiologic Risk Score

Carri W. Chan, Columbia University

Abstract: Unplanned transfers of patients from general medical-surgical wards to the Intensive Care Unit (ICU) can occur due to unexpected patient deterioration. Such patients tend to have higher mortality rates and longer lengths-of-stay than direct admissions to the ICU. As such, the medical community has invested substantial efforts in the development of patient risk scores with the intent to identify patients at risk of deterioration. In this work, we consider how one such risk score could be used to trigger proactive transfers to the ICU. We utilize a retrospective dataset from 21 Kaiser Permanente Northern California hospitals to estimate the potential benefit of transferring patients to the ICU at various levels of patient risk of deterioration. In order to reduce the sensitivity of our findings to key identification and modeling assumptions, we use a combination of multivariate matching and instrumental variable approaches. We also study the impact of parameter uncertainty that arises when models are estimated from real world data and provide recommendations that are robust to unavoidable parameter misspecification and estimation errors. We find that proactively transferring the most severe patients could reduce mortality rates and lengths-of-stay without increasing other adverse events; however, proactive transfers should be used judiciously as being too aggressive could increase ICU congestion and degrade quality of care.

Papers:

1. J. Grand-Clement, C. W. Chan, V. Goyal, G. Escobar, Robustness of proactive ICU transfer policies
2. W. Hu, C. W. Chan, J. R. Zubizarreta, G. J. Escobar, An Examination of Early Transfers to the ICU Based on a Physiologic Risk Score, MSOM, 2018.

Sparse Flexible Design for Radiation Therapy: A Machine Learning Approach.

Timothy Chan, University of Toronto

Abstract: For a general production network, state-of-the-art methods for constructing sparse flexible designs are heuristic in nature, typically computing a proxy for the quality of unseen networks and using that estimate in a greedy manner to modify a current design. This paper develops two machine learning-based approaches to constructing sparse flexible designs that leverage a neural network to accurately and quickly predict the performance of large numbers of candidate designs. We demonstrate that our heuristics are competitive with existing approaches and produce high-quality solutions for both balanced and unbalanced networks. Finally, we introduce a novel application of process flexibility in healthcare operations to demonstrate the effectiveness of our approach in a large numerical case study. We study the flexibility of linear accelerators that deliver radiation to treat various types of cancer. We demonstrate how clinical constraints can be easily absorbed into the machine learning subroutine and how our sparse flexible treatment networks meet or beat the performance of those designed by state-of-the-art methods.

Watch a video of the presentation →
Download a copy of the presentation slides →

Exploring Regional Variation and Survival in Colon Cancer Pathway Concordance.

Luciano Ieraci and Maria Eberg, Ontario Health (Cancer Care Ontario)

Abstract: Clinical pathways have been defined as structured multidisciplinary plans outlining care for patients with specific health conditions, such as cancer. Pathways are increasingly used in clinical practice and can improve some aspects of care. Measuring concordance of actual care with cancer pathway recommendations (i.e. reference pathways) could identify opportunities to improve care further. However, population-level pathway concordance measures for the entire cancer care trajectory are lacking. Investigations of pathway concordance with specific patient outcomes is also absent in published literature. In the current analyses, we have developed both a characterization of pathway concordance and found its association with overall patient survival.

Measuring the difference between reference and observed pathways was facilitated by using pre-existing methods on string similarity. Each pathway can be represented as a sequence of events and comparisons made between reference pathway sequences and observed sequences according to the events actually received by patients. Levenshtein distance is one such comparison algorithm. We developed a measure based on the Levenshtein algorithm to quantify concordance with a simplified colon cancer pathway map. To develop and validate it, we used a cohort of stage II and III colon cancer patients diagnosed from 2012-2016 in Ontario, Canada. We investigated association between concordance and patient survival at the population level, and compared the measure’s discrimination between survivors and decedents. Concordance scores were significantly associated with patient survival and discriminated between more patient pathways. We also examined the variation in concordance scores across health regions and characterized the differences by comparing the prevalence of specific patient subgroups across regions.

Overall, our Levenshtein-based measure incorporated differences between actual care and simplified pathways, was strongly associated with patient survival, and demonstrated good patient discrimination. The developed measure is a potentially valuable tool for health system performance monitoring and quality improvement.

Watch a video of the presentation →
Download a copy of the presentation slides →

Personalized Treatment Adherence Support Strategies for Tuberculosis Patients in Kenya

Jonas Oddur Jonasson, MIT

Abstract: Treatment adherence among tuberculosis patients is not only critical for the cure of the individual patients but also for public health, as it minimizes the risk of transmission and the development of drug resistant strains of the disease. Many studies have examined the impact of various treatment adherence support (TAS) programs on patient outcomes and completion rates, with very mixed results. We use data from a completed randomized-controlled trial, which evaluated the impact of a TAS program that requires daily patient engagement, to develop personalized enrollment and outreach strategies. First, we generate personalized pre-enrollment predictions of the impact of a TAS program on treatment outcomes. We then develop an enrollment strategy as a function of the relative cost of providing the service and the established benefits of individual treatment completion. Second, we generate a personalized risk prediction algorithm for patients already enrolled. Specifically, we consider (A) the long-term risk of incomplete treatment and (B) the short-term risk of reduced engagement with the platform. Our analysis demonstrates the value of patient engagement information for prioritizing treatment adherence support.

Watch a video of the presentation →
Download a copy of the slides →

Hypertension Management: A Value of Information Perspective.

Manaf Zargoush, Degroote School of Business, McMaster University

Abstract: We present an analytical framework that incorporates several features of the hypertension management process. Fundamental to the framework is the modelling of measurement noise as well as the individual short-term and long-term blood pressure variabilities. The physician in charge of care sees the patient with a certain frequency and prescribes medication based on blood pressure measurements. There are a few technologies for measuring blood pressure ranging from the traditional method of arm readings to applanation tonometry, a more invasive technique providing readings much closer to the heart. We conceptualize that the physician forms a belief about the patient’s true blood pressure state at every visit. This leads to two components operating concurrently in the proposed framework. First is a learning module representing how clinical judgments about the patient’s blood pressure status, based on the noisy and stochastic observations, are made sequentially. Here, we adopt a learning strategy to incorporate the role of cognitive bias that is a consequence of the uncertainties around blood pressure observations. The second component is the optimization module representing how the treatment decisions (i.e., medication prescriptions) are optimized under learning. The optimal course of actions involves minimizing the cardiovascular disease risks and the medication side-effects. The care delivery is tailored to the patient’s individual case mix variables. Our framework is highly consistent with the current medical practices, boosting the validity of our models and the expected buy-in from practitioners to the insights we provide. To this end, we calibrate our models with the data collected in hypertension clinic of Montreal General Hospital supplemented with parameter estimates from the clinical literature.

Watch a video of the presentation →
Download a copy of the presentation slides →

Big Data - Local Impact - Healthy Lives

Laura Derksen, University of Toronto

Abstract: The emergence of electronic medical records (EMR) in low resource settings has changed the way care is provided on the ground, and generates many new opportunities for research. I will present several ongoing projects studying, and making use of, Malawi's new EMR system. Malawi is a low resource country in southern Africa, with a serious health burden, especially related to HIV/AIDS, malaria, infant health, and increasingly, non-communicable diseases. First, I will present preliminary results, using a difference-in-difference strategy to estimate the impact of EMR on HIV patient outcomes, and explore improvements in patient tracking as a mechanism. Second, I will present work by PhD student Jessica Gallant using machine learning and EMR data to predict patient outcomes based on health provider identities and characteristics. Third, I will discuss plans for a randomized controlled trial to investigate the effect of data collection and data privacy policy on patient demand for health services. Data privacy is a key concern, especially for HIV/AIDS patients, and the collection of electronic medical data may directly impact a person’s desire to seek care.

Watch a video of the presentation →
Download a copy of the presentation slides →

Artificial Intelligence in Healthcare? Evidence from online job postings.

Avi Goldfarb, University of Toronto

Abstract: This paper documents a puzzle. Despite the numerous popular press discussions of artificial intelligence (AI) in healthcare, there has been relatively little adoption. Using data from Burning Glass Technologies on millions of online job postings, we find that AI adoption in healthcare remains substantially lower than in most other industries, and that under 3% of the hospitals in our data posted any jobs requiring AI skills from 2015-2018. The low adoption rates mean any statistical analysis is limited. Nevertheless, the adoption we do observe shows that larger hospitals, larger counties, and integrated salary model hospitals are more likely to adopt.

Watch a video of the presentation →
Download a copy of the presentation slides →