programming code on a computer monitor
Photo by Markus Spiske on Unsplash.

The Problem

Preventable surgical complications, which accounts for 15% of all surgical procedures performed, continue to have significant negative impact on patients’ quality of life, surgeons’ reimbursement, and health systems’ cost of care. However, current risk assessment methods practiced by surgeons are almost entirely subjective. Even objective tools such as the American College of Surgeon’s Risk Calculator creates too much burden for the surgeon to use on every patient. Published analyses of Medicare costs of inpatient surgery demonstrate that payments for the index hospitalization usually account for the majority of increased payments (Birkmeyer, Annals of Surgery, 2012). Yet, in spite of great progress in data analytic techniques, the “risk space” for post-operative outcomes is poorly explored. Unmanaged variability in post-operative outcomes leads to preventable morbidity and mortality for the patient, economic cost to society, with complications potentially multiplying the cost of a procedure by a factor of five (Vonlanthen, Annals of Surgery, 2011).

Our Solution

We propose that machine learning—class discovery and predictive modeling techniques—provide the analytic toolset for understanding and managing post-operative outcomes by allowing practitioners to better understand variables they can and cannot control and providing a predictive and quantitative basis for optimizing surgical and post-surgical management. The solution created a comprehensive computing and risk stratification platform for:

  1. Predicting risk of surgical complications
  2. Tailoring interventions to the individual patient, and
  3. Improving surgical outcomes through best practice measures.

Essential for implementing such management is a platform that enables:

  1. Rigorously analyzing the many variables that contribute to surgical outcomes
  2. Melding these variables into risk stratification models
  3. Deploying these models with provider-friendly interfaces
  4. Designing data-driven interventions

Such a platform represents the key to effectively using research and operational data to scalably enhance the effectiveness and efficiency of peri-operative interventions throughout the health system. CALYPSO (Clinical Analytic & Learning Platform in Surgical Outcomes) is a platform for employing machine learning strategies to identify higher-order relationships in local data and national databases such as NSQIP and ERAS, that facilitates transfer learning between such datasets, and provides the means to operationalize insights via “predictive models as a service” within the Health System.

Impact

The team developed a highly predictive model using novel methodology, leveraging both national NSQIP (National Surgery Quality Improvement Program) and Duke surgical data. This model was built off of 4 million national records. The model also predicts on 11 clinical outcomes (wound infections, dvt, etc.). The team designed and created an engaging user interface to deliver prediction results and interventions to the end-users.

The team also completed a companion pilot study (funded through the DIHI Innovation Jam by the Department of Surgery as well as additional funding from Kevin Sowers) for measuring patients’ 30-day complication outcomes while using our platform. In this study we rounded on 200 patients and data is currently being analyzed. The team is working with Duke Health Technology Systems and Epic on product integration.

  • The innovative risk prediction model has been deployed on daily rounding of surgical patients
  • Follow-on funding by Department of Surgery and Kevin Sowers (DUH). »An IDF was filed with the Office of Licensing and Ventures.
  • A company was created, KelaHealth, which is in the final stages of negotiation with OLV on licensure. Angel investors engaged.
  • Academic: 2 oral presentations at national conferences, 1 publication, and awaiting second publication.

More Projects