Surgical Epidemiology: Prediction and Causation in Clinical Datasets

Abstract

Many important questions on the comparative effectiveness of surgical interventions may be difficult to answer with randomized control trials due to ethical, feasibility, or logistical concerns. While the field of causal inference has advanced our ability to interrogate causal questions in medicine using observational data, recent methodological frameworks such as target trial emulation are not commonly utilized in surgical literature. The use of these methods in surgical epidemiology is complicated by the fact that many important indications for surgical interventions, which are usually confounders for the effect of interest, are not captured in large, automated datasets. Using electronic medical record data, including data manually collected from operative and visit notes, we interrogated questions arising from the fields of orthopedic surgery (Chapters 1 and 2) and obstetric surgery (Chapter 3) using both prediction (Chapter 1) and causal inference (Chapters 2 and 3) methodologies. These studies are united by their use of clinical datasets to answer questions involving relatively uncommon surgical interventions.

In Chapter 1, we identified two cohorts of surgically treated scaphoid non-unions from the United States and Japan (N=70) and used computed tomography (CT) scans to measure the bone density of proximal and distal fracture fragments relative to the triquetrum. In using a relative measure, we aimed to isolate within-subject differences in scaphoid bone density from between-subject differences in baseline bone density due to age, sex, nutrition, etc. We assessed the predictive characteristics of this measure for post-operative healing within 3 months. The results may be used to counsel patients considering surgery on the likelihood of post-operative healing.

In Chapter 2, we assessed the effect of an orthopedic surgical intervention with strong clinical equipoise and weak a priori hypothesized confounding. We specified a target trial to estimate the effect of anterior transposition versus in situ release of the ulnar nerve on the 2-year risk of reoperation for ulnar neuropathy within a retrospective cohort of surgically treated distal humerus fractures. Although prior randomized control trials have been conducted to assess the effect of ulnar nerve transposition on post-operative ulnar neuropathy, neuropathy is a subjective outcome that may be transient in nature. This is the first study to describe the epidemiology of reoperation for ulnar neuropathy among patients with distal humerus fracture and to assess the effect of ulnar nerve transposition on this outcome.

In Chapter 3, we applied a similar methodology to assess the effect of an obstetric surgical intervention with strong a priori hypothesized confounding. We additionally utilized an alternative methodology to assess the population-level effect of more pragmatic treatment strategies to address a lack of equipoise for subsets of patients included in static treatment strategies. Transabdominal cervical cerclage is an abdominal surgical procedure to prevent spontaneous preterm birth. The first and only randomized trial on transabdominal cerclage found a strongly protective effect for spontaneous preterm delivery before 32 weeks of gestation compared with transvaginal cervical cerclage. However, this trial was conducted in a high-risk obstetric population and the effect size may not generalize to a patient population with fewer risk factors for recurrent spontaneous preterm birth. We specified a target trial to estimate the effect of transabdominal cervical cerclage versus transvaginal cerclage on the risk of preterm delivery before 34 weeks within a lower risk obstetric population compared with the trial population. The results of this study may be used to inform decisions about whether transabdominal cerclage should be offered to lower risk patients.