An explainable long short-term memory network for surgical site infection identification

BACKGROUND

Currently, surgical site infection surveillance relies on labor-intensive manual chart review. Recently suggested solutions involve machine learning to identify surgical site infections directly from the medical record. Deep learning is a form of machine learning that has historically performed better than traditional methods while being harder to interpret. We propose a deep learning model, a long short-term memory network, for the identification of surgical site infection from the medical record with an attention layer for explainability.

METHODS

We retrieved structured data and clinical notes from the University of Utah Health System's electronic health care record for operative events randomly selected for manual chart review from January 2016 to June 2021. Surgical site infection occurring within 30 days of surgery was determined according to the National Surgical Quality Improvement Program definition. We trained the long short-term memory model along with traditional machine learning models for comparison. We calculated several performance metrics from a holdout test set and performed additional analyses to understand the performance of the long short-term memory, including an explainability analysis.

RESULTS

Surgical site infection was present in 4.7% of the total 9,185 operative events. The area under the receiver operating characteristic curve and sensitivity of the long short-term memory was higher (area under the receiver operating characteristic curve: 0.954, sensitivity: 0.920) compared to the top traditional model (area under the receiver operating characteristic curve: 0.937, sensitivity: 0.736). The top 5 features of the long short-term memory included 2 procedure codes and 3 laboratory values.

CONCLUSION

Surgical site infection surveillance is vital for the reduction of surgical site infection rates. Our explainable long short-term memory achieved a comparable area under the receiver operating characteristic curve and greater sensitivity when compared to traditional machine learning methods. With explainable deep learning, automated surgical site infection surveillance could replace burdensome manual chart review processes.

An artificial intelligence approach for investigating multifactorial pain-related features of endometriosis

Endometriosis is a debilitating, chronic disease that is estimated to affect 11% of reproductive-age women. Diagnosis of endometriosis is difficult with diagnostic delays of up to 12 years reported. These delays can negatively impact health and quality of life. Vague, nonspecific symptoms, like pain, with multiple differential diagnoses contribute to the difficulty of diagnosis. By investigating previously imprecise symptoms of pain, we sought to clarify distinct pain symptoms indicative of endometriosis, using an artificial intelligence-based approach. We used data from 473 women undergoing laparoscopy or laparotomy for a variety of surgical indications. Multiple anatomical pain locations were clustered based on the associations across samples to increase the power in the probability calculations. A Bayesian network was developed using pain-related features, subfertility, and diagnoses. Univariable and multivariable analyses were performed by querying the network for the relative risk of a postoperative diagnosis, given the presence of different symptoms. Performance and sensitivity analyses demonstrated the advantages of Bayesian network analysis over traditional statistical techniques. Clustering grouped the 155 anatomical sites of pain into 15 pain locations. After pruning, the final Bayesian network included 18 nodes. The presence of any pain-related feature increased the relative risk of endometriosis (p-value < 0.001). The constellation of chronic pelvic pain, subfertility, and dyspareunia resulted in the greatest increase in the relative risk of endometriosis. The performance and sensitivity analyses demonstrated that the Bayesian network could identify and analyze more significant associations with endometriosis than traditional statistical techniques. Pelvic pain, frequently associated with endometriosis, is a common and vague symptom. Our Bayesian network for the study of pain-related features of endometriosis revealed specific pain locations and pain types that potentially forecast the diagnosis of endometriosis.

Developing an LSTM Model to Identify Surgical Site Infections using Electronic Healthcare Records

Recently, hospitals and healthcare providers have made efforts to reduce surgical site infections as they are a major cause of surgical complications, a prominent reason for hospital readmission, and associated with significantly increased healthcare costs. Traditional surveillance methods for SSI rely on manual chart review, which can be laborious and costly. To assist the chart review process, we developed a long short-term memory (LSTM) model using structured electronic health record data to identify SSI. The top LSTM model resulted in an average precision (AP) of 0.570 [95% CI 0.567, 0.573] and area under the receiver operating characteristic curve (AUROC) of 0.905 [95% CI 0.904, 0.906] compared to the top traditional machine learning model, a random forest, which achieved 0.552 [95% CI 0.549, 0.555] AP and 0.899 [95% CI 0.898, 0.900] AUROC. Our LSTM model represents a step toward automated surveillance of SSIs, a critical component of quality improvement mechanisms.

Standard Vocabularies to Improve Machine Learning Model Transferability With Electronic Health Record Data: Retrospective Cohort Study Using Health Care-Associated Infection

BACKGROUND

With the widespread adoption of electronic healthcare records (EHRs) by US hospitals, there is an opportunity to leverage this data for the development of predictive algorithms to improve clinical care. A key barrier in model development and implementation includes the external validation of model discrimination, which is rare and often results in worse performance. One reason why machine learning models are not externally generalizable is data heterogeneity. A potential solution to address the substantial data heterogeneity between health care systems is to use standard vocabularies to map EHR data elements. The advantage of these vocabularies is a hierarchical relationship between elements, which allows the aggregation of specific clinical features to more general grouped concepts.

OBJECTIVE

This study aimed to evaluate grouping EHR data using standard vocabularies to improve the transferability of machine learning models for the detection of postoperative health care-associated infections across institutions with different EHR systems.

METHODS

Patients who underwent surgery from the University of Utah Health and Intermountain Healthcare from July 2014 to August 2017 with complete follow-up data were included. The primary outcome was a health care-associated infection within 30 days of the procedure. EHR data from 0-30 days after the operation were mapped to standard vocabularies and grouped using the hierarchical relationships of the vocabularies. Model performance was measured using the area under the receiver operating characteristic curve (AUC) and F₁-score in internal and external validations. To evaluate model transferability, a difference-in-difference metric was defined as the difference in performance drop between internal and external validations for the baseline and grouped models.

RESULTS

A total of 5775 patients from the University of Utah and 15,434 patients from Intermountain Healthcare were included. The prevalence of selected outcomes was from 4.9% (761/15,434) to 5% (291/5775) for surgical site infections, from 0.8% (44/5775) to 1.1% (171/15,434) for pneumonia, from 2.6% (400/15,434) to 3% (175/5775) for sepsis, and from 0.8% (125/15,434) to 0.9% (50/5775) for urinary tract infections. In all outcomes, the grouping of data using standard vocabularies resulted in a reduced drop in AUC and F₁-score in external validation compared to baseline features (all P<.001, except urinary tract infection AUC: P=.002). The difference-in-difference metrics ranged from 0.005 to 0.248 for AUC and from 0.075 to 0.216 for F₁-score.

CONCLUSIONS

We demonstrated that grouping machine learning model features based on standard vocabularies improved model transferability between data sets across 2 institutions. Improving model transferability using standard vocabularies has the potential to improve the generalization of clinical prediction models across the health care system.

Lung retrieval from non-heart beating cadavers with the use of a rat lung transplant model

BACKGROUND

Lungs retrieved from cadavers after death and circulatory arrest may alleviate the critical shortage of lungs for transplant. We report a rat lung transplantation model that allows serial measurement of arterial blood gases after left single lung transplantation from non-heart beating donors.

METHODS

Twelve Sprague-Dawley rats underwent left lung transplantation with a vascular cuff technique. Donor rats were anesthetized with intraperitoneal injection of pentobarbital, heparinized, intubated via tracheotomy, and then killed with pentobarbital. Lungs were retrieved immediately or after 2 hours of oxygen ventilation after death (tidal volume 1 mL/100 g, rate 40/min FIO2 = 1.0, positive end-expiratory pressure 5 cm H2O). Recipient rats were anesthetized, intubated, and ventilated. The carotid artery and jugular vein were cannulated for arterial blood gases and infusion of Ringer's lactate (4 mL/h). Anesthesia was maintained with halothane 0.2%, and recipient arterial blood gases were measured at 4 and 6 hours after lung transplantation after snaring the right pulmonary artery for 5 minutes. Animals were put to death 6 hours after lung transplantation, and portions of transplanted lungs were frozen in liquid nitrogen and assayed for wet/dry ratio, myeloperoxidase as a measure of neutrophil infiltration, and conjugated dienes as a measure of free radical-mediated lipid peroxidation.

RESULTS

Arterial PO2 and wet/dry ratio were not significantly different in recipients of non-heart beating donor lungs retrieved immediately after death or after 2 hours of oxygen ventilation. Significant neutrophil infiltration was observed in recipients of non-heart beating donor lungs retrieved 2 hours after death from oxygen-ventilated donors.

CONCLUSIONS

Strategies to ameliorate reperfusion injury may allow for successful lung transplantation from non-heart beating donors.

Blunt tracheobronchial injuries: treatment and outcomes

BACKGROUND

Tracheobronchial injury is a recognized, yet uncommon, result of blunt trauma to the thorax. Often the diagnosis and treatment are delayed, resulting in attempted surgical repair months or even years after the injury. This report is an extensive review of the literature on tracheobronchial ruptures that examines outcomes and their association with the time from injury to diagnosis.

METHODS

We reviewed all patients with blunt tracheobronchial injuries published in the literature to determine the anatomic location of the injury, mechanism of the injury, time until diagnosis and treatment, and outcome. Only patients with blunt intrathoracic tracheobronchial traumas were included.

RESULTS

We identified 265 patients reported between 1873 and 1996. Motor vehicle accidents were the most frequent mechanism of injury (59%). The overall mortality among reported patients has declined from 36% before 1950 to 9% since 1970. The injury occurred within 2 cm of the carina in 76% of patients, and 43% occurred within the first 2 cm of the right main bronchus. The proximity of the injury to the carina had no detectable effect on mortality. Injuries on the right side were treated sooner but were associated with a higher mortality than left-sided injuries. No association was detected between delay in treatment and successful repair of the injury; ninety percent of patients undergoing treatment more than 1 year after injury were repaired successfully.

CONCLUSIONS

This review of patients with blunt tracheobronchial injuries represents the largest cohort studied to date. These data suggest an ability to repair tracheobronchial injuries successfully many months after they occur. We are also able to assess the mortality associated with the location and side of injury, examine the time from injury until diagnosis and treatment, and evaluate treatment outcome.

Spontaneous hemopneumothorax in women

Spontaneous hemopneumothorax is uncommon, especially among women. We report a case of spontaneous hemopneumothorax in a 19-year-old woman and review seven other cases of spontaneous hemopneumothorax in women that have been reported in the English language.