Career Trajectory After General Surgery Residency: Do Academic Program Graduates Pursue Academic Surgery?

OBJECTIVE: Determine the proportion of contemporary US academic general surgery residency program graduates who pursue academic careers and identify factors associated with pursuing academic careers. SUMMARY BACKGROUND DATA: Many academic residency programs aim to cultivate academic surgeons, yet the proportion of contemporary graduates who choose academic careers is unclear. The potential determinants that affect graduates' decisions to pursue academic careers remain underexplored. METHODS: We collected program and individual-level data on 2015 and 2018 graduates across 96 US academic general surgery residency programs using public resources. We defined those pursuing academic careers as faculty within US allopathic medical school-affiliated surgery departments who published two or more peer-reviewed publications as the first or senior author between 2020-2021. After variable selection using sample splitting LASSO regression, multivariable regression evaluated association with pursuing academic careers among all graduates, and graduates of top-20 residency programs. Secondary analysis using multivariable ordinal regression explored factors associated with high research productivity during early faculty years. RESULTS: Among 992 graduates, 166 (17%) were pursuing academic careers according to our definition. Graduating from a top-20 ranked residency program (OR[95%CI]: 2.34[1.40-3.88]), working with a longitudinal research mentor during residency (OR[95%CI]: 2.21[1.24-3.95]), holding an advanced degree (OR[95%CI]: 2.20[1.19-3.99]), and the number of peer-reviewed publications during residency as first or senior author (OR[95%CI]: 1.13[1.07-1.20]) were associated with pursuing an academic surgery career, while the number of peer-reviewed publications before residency was not (OR[95%CI]: 1.08[0.99-1.20]). Among top 20 program graduates, working with a longitudinal research mentor during residency (OR[95%CI]: 0.95[0.43-2.09]) was not associated with pursuing an academic surgery career. The number of peer-reviewed publications during residency as first or senior author was the only variable associated with higher productivity during early faculty years (OR[95%CI]: 1.12[1.07-1.18]). CONCLUSIONS: Our findings suggest programs that aim to graduate academic surgeons may benefit from ensuring trainees receive infrastructural support and demonstrate sustained commitment to research throughout residency. Our results should be interpreted cautiously as the impact of unmeasured confounders is unclear.

Explainable Machine Learning to Bring Database to the Bedside: Development and Validation of the TROUT (Trauma fRailty OUTcomes) Index, a Point-of-Care Tool to Prognosticate Outcomes After Traumatic Injury Based on Frailty.

OBJECTIVE: Exemplify an explainable machine learning framework to bring database to the bedside; develop and validate a point-of-care frailty assessment tool to prognosticate outcomes after injury. BACKGROUND: A geriatric trauma frailty index that captures only baseline conditions, is readily-implementable, and validated nationwide remains underexplored. We hypothesized Trauma fRailty OUTcomes (TROUT) Index could prognosticate major adverse outcomes with minimal implementation barriers. METHODS: We developed TROUT index according to Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis guidelines. Using nationwide US admission encounters of patients aged ≥65 years (2016-2017; 10% development, 90% validation cohorts), unsupervised and supervised machine learning algorithms identified baseline conditions that contribute most to adverse outcomes. These conditions were aggregated into TROUT Index scores (0-100) that delineate 3 frailty risk strata. After associative [between frailty risk strata and outcomes, adjusted for age, sex, and injury severity (as effect modifier)] and calibration analysis, we designed a mobile application to facilitate point-of-care implementation. RESULTS: Our study population comprised 1.6 million survey-weighted admission encounters. Fourteen baseline conditions and 1 mechanism of injury constituted the TROUT Index. Among the validation cohort, increasing frailty risk (low=reference group, moderate, high) was associated with stepwise increased adjusted odds of mortality {odds ratio [OR] [95% confidence interval (CI)]: 2.6 [2.4-2.8], 4.3 [4.0-4.7]}, prolonged hospitalization [OR (95% CI)]: 1.4 (1.4-1.5), 1.8 (1.8-1.9)], disposition to a facility [OR (95% CI): 1.49 (1.4-1.5), 1.8 (1.7-1.8)], and mechanical ventilation [OR (95% CI): 2.3 (1.9-2.7), 3.6 (3.0-4.5)]. Calibration analysis found positive correlations between higher TROUT Index scores and all adverse outcomes. We built a mobile application ("TROUT Index") and shared code publicly. CONCLUSION: The TROUT Index is an interpretable, point-of-care tool to quantify and integrate frailty within clinical decision-making among injured patients. The TROUT Index is not a stand-alone tool to predict outcomes after injury; our tool should be considered in conjunction with injury pattern, clinical management, and within institution-specific workflows. A practical mobile application and publicly available code can facilitate future implementation and external validation studies.

TraumaICDBERT, A Natural Language Processing Algorithm to Extract Injury ICD-10 Diagnosis Code from Free Text.

OBJECTIVE: To develop and validate TraumaICDBERT, a natural language processing algorithm to predict injury ICD-10 diagnosis codes from trauma tertiary survey notes. SUMMARY BACKGROUND DATA: The adoption of ICD-10 diagnosis codes in clinical settings for injury prediction is hindered by the lack of real-time availability. Existing natural language processing algorithms have limitations in accurately predicting injury ICD-10 diagnosis codes. METHODS: Trauma tertiary survey notes from hospital encounters of adults between January 2016 and June 2021 were used to develop and validate TraumaICDBERT, an algorithm based on BioLinkBERT. The performance of TraumaICDBERT was compared to Amazon Web Services Comprehend Medical, an existing natural language processing tool. RESULTS: A dataset of 3,478 tertiary survey notes with 15,762 4-character injury ICD-10 diagnosis codes was analyzed. TraumaICDBERT outperformed Amazon Web Services Comprehend Medical across all evaluated metrics. On average, each tertiary survey note was associated with 3.8 (standard deviation: 2.9) trauma registrar-extracted 4-character injury ICD-10 diagnosis codes. CONCLUSIONS: TraumaICDBERT demonstrates promising initial performance in predicting injury ICD-10 diagnosis codes from trauma tertiary survey notes, potentially facilitating the adoption of downstream prediction tools in clinical settings.

Artificial Intelligence-enabled Decision Support in Surgery: State-of-the-art and Future Directions.

OBJECTIVE: To summarize state-of-the-art artificial intelligence-enabled decision support in surgery and to quantify deficiencies in scientific rigor and reporting. BACKGROUND: To positively affect surgical care, decision-support models must exceed current reporting guideline requirements by performing external and real-time validation, enrolling adequate sample sizes, reporting model precision, assessing performance across vulnerable populations, and achieving clinical implementation; the degree to which published models meet these criteria is unknown. METHODS: Embase, PubMed, and MEDLINE databases were searched from their inception to September 21, 2022 for articles describing artificial intelligence-enabled decision support in surgery that uses preoperative or intraoperative data elements to predict complications within 90 days of surgery. Scientific rigor and reporting criteria were assessed and reported according to Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews guidelines. RESULTS: Sample size ranged from 163-2,882,526, with 8/36 articles (22.2%) featuring sample sizes of less than 2000; 7 of these 8 articles (87.5%) had below-average (<0.83) area under the receiver operating characteristic or accuracy. Overall, 29 articles (80.6%) performed internal validation only, 5 (13.8%) performed external validation, and 2 (5.6%) performed real-time validation. Twenty-three articles (63.9%) reported precision. No articles reported performance across sociodemographic categories. Thirteen articles (36.1%) presented a framework that could be used for clinical implementation; none assessed clinical implementation efficacy. CONCLUSIONS: Artificial intelligence-enabled decision support in surgery is limited by reliance on internal validation, small sample sizes that risk overfitting and sacrifice predictive performance, and failure to report confidence intervals, precision, equity analyses, and clinical implementation. Researchers should strive to improve scientific quality.

Systematic Review and Meta-Analysis of Hardware Failure in Surgical Stabilization of Rib Fractures: Who, What, When, Where, and Why?

BACKGROUND: Surgical stabilization of rib fractures (SSRF) is increasingly used to reduce pulmonary complications and death among patients with rib fractures. However, the five Ws of hardware failure -who, what, when, where, and why- remains unclear. We aimed to synthesize available evidence on the five Ws and outline future research agenda for mitigating hardware failure. METHODS: Experimental and observational studies published between 2009 and 2020 evaluating adults undergoing SSRF for traumatic rib fractures underwent evidence synthesis. We performed random effects meta-analysis of cohort/consecutive case studies. We calculated pooled prevalence of SSRF hardware failures using Freeman-Tukey double arcsine transformation and assessed study heterogeneity using DerSimonian-Laird estimation. We performed meta-regression with rib fracture acuity (acute or chronic) and hardware type (metal plate or not metal plate) as moderators. RESULTS: Twenty-nine studies underwent qualitative synthesis and 24 studies (2404 SSRF patients) underwent quantitative synthesis. Pooled prevalence of hardware failure was 4(3-7)%. Meta-regression showed fracture acuity was a significant moderator (P = 0.002) of hardware failure but hardware type was not (P = 0.23). Approximately 60% of patients underwent hardware removal after hardware failure. Mechanical failures were the most common type of hardware failure, followed by hardware infections, pain/discomfort, and non-union. Timing of hardware failure after surgery was highly variable, but 87% of failures occurred after initial hospitalization. Mechanical failures was attributed to technical shortcomings (i.e. short plate length) or excessive force on the thoracic cavity. CONCLUSIONS: SSRF hardware failure is an uncommon complication. Not all hardware failures are consequential, but insufficient individual patient data precluded characterizing where and why hardware failures occur. Minimizing SSRF hardware failure requires concerted research agenda to expand on the paucity of existing evidence.

Prospectively Assigned AAST Grade versus Modified Hinchey Class and Acute Diverticulitis Outcomes.

BACKGROUND: The American Association for the Surgery of Trauma (AAST) recently developed a classification system to standardize outcomes analyses for several emergency general surgery conditions. To highlight this system's full potential, we conducted a study integrating prospective AAST grade assignment within the electronic medical record. METHODS: Our institution integrated AAST grade assignment into our clinical workflow in July 2018. Patients with acute diverticulitis were prospectively assigned AAST grades and modified Hinchey classes at the time of surgical consultation. Support vector machine-a machine learning algorithm attuned for small sample sizes-was used to compare the associations between the two classification systems and decision to operate and incidence of complications. RESULTS: 67 patients were included (median age of 62 y, 40% male) for analysis. The decision for operative management, hospital length of stay, intensive care unit admission, and intensive care unit length of stay were associated with both increasing AAST grade and increasing modified Hinchey class (all P < 0.001). AAST grade additionally showed a correlation with complication severity (P = 0.02). Compared with modified Hinchey class, AAST grade better predicted decision to operate (88.2% versus 82.4%). CONCLUSIONS: This study showed the feasibility of electronic medical record integration to support the full potential of AAST classification system's utility as a clinical decision-making tool. Prospectively assigned AAST grade may be an accurate and pragmatic method to find associations with outcomes, yet validation requires further study.

Early National Landscape of Surgical Stabilization of Sternal Fractures.

BACKGROUND: Operative management of chest wall injuries aims to restore respiratory mechanics and mitigate pulmonary complications. Extensive studies support surgical stabilization of rib fractures (SSRF) for select patients, but role for surgical stabilization of sternal fractures (SSSF) remains unclear. We aimed to understand national prevalence of SSSF and compare outcomes after surgical stabilization and non-operative management of sternal fractures. METHODS: We retrospectively analyzed adult patients (age ≥ 18 years) admitted with sternal fractures after blunt trauma using the 2016 National Trauma Data Bank. We compared odds of inpatient mortality, pneumonia, and respiratory failure for propensity score matched patients (4:1) who underwent non-operative management vs SSSF. We characterized subgroup of patients with concurrent rib and sternal fractures who underwent concomitant SSRF-SSSF. RESULTS: We identified 14,760 encounters of adults admitted with sternal fractures; 270 (1.8%) underwent SSSF. Compared to matched patients who underwent non-operative management, patients who underwent SSSF had lower odds of mortality (OR [95%CI]: 0.19 [0.06-0.62], p = 0.006). Adjusted for trauma center level, Mantel-Haenszel mortality odds remained lower for patients who underwent SSSF. Odds of pneumonia and respiratory failure were similar between matched groups. Among 46% of patients who had concomitant rib fractures, 0.3% (n = 18) underwent concurrent SSRF-SSSF and these patients survived hospitalization without pneumonia or respiratory failure. CONCLUSION: A vast majority of patients who suffer sternal fractures undergo non-operative management. Potential mortality benefit of SSSF and concurrent SSRF-SSSF's role for commonly concomitant rib and sternal fractures deserve further study. Our preliminary findings call for delineating heterogeneity of sternal fractures and establishing consensus SSSF indications.

The burden of readmissions after rib fractures among older adults.

BACKGROUND: The index hospitalization morbidity and mortality of rib fractures among older adults (aged ≥65 years) is well-known, yet the burden and risks for readmissions after rib fractures in this vulnerable population remain understudied. We aimed to characterize the burdens and etiologies associated with 3-month readmissions among older adults who suffer rib fractures. We hypothesized that readmissions would be common and associated with modifiable etiologies. METHODS: This survey-weighted retrospective study using the 2017 and 2019 National Readmissions Database evaluated adults aged ≥65 years hospitalized with multiple rib fractures and without major extrathoracic injuries. The main outcome was the proportion of patients experiencing all-cause 3-month readmissions. We assessed the 5 leading principal readmission diagnoses overall and delineated them by index hospitalization discharge disposition (home or facility). Sensitivity analysis using clinical classification categories characterized readmissions that could reasonably represent rib fracture-related sequelae. RESULTS: In 2017, 25,092 patients met the inclusion criteria, with 20% (N = 4,894) experiencing 3-month readmissions. Six percent of patients did not survive their readmission. The 5 leading principal readmission diagnoses were sepsis (many associated with secondary diagnoses of pneumonia [41%] or urinary tract infections [41%]), hypertensive heart/kidney disease, hemothorax, pneumonia, and respiratory failure. In 2019, a comparable 3-month readmission rate of 23% and identical 5 leading diagnoses were found. Principal readmission diagnosis of hemothorax was associated with the shortest time to readmission (median [interquartile range]:9 [5-23] days). Among patients discharged home after index hospitalization, pleural effusion-possibly representing mischaracterized hemothorax-was among the leading principal readmission diagnoses. Some patients readmitted with a principal diagnosis of hemothorax or pleural effusion had these diagnoses at index hospitalization; a lower proportion of these patients underwent pleural fluid intervention during index hospitalization compared with readmission. On sensitivity analysis, 30% of 3-month readmissions were associated with principal diagnoses suggesting rib fracture-related sequelae. CONCLUSION: Readmissions are not infrequent among older adults who suffer rib fractures, even in the absence of major extrathoracic injuries. Future studies should better characterize how specific complications associated with readmissions, such as pneumonia, urinary tract infections, and delayed hemothoraces, could be mitigated.

Practical guide to building machine learning-based clinical prediction models using imbalanced datasets.

Clinical prediction models often aim to predict rare, high-risk events, but building such models requires robust understanding of imbalance datasets and their unique study design considerations. This practical guide highlights foundational prediction model principles for surgeon-data scientists and readers who encounter clinical prediction models, from feature engineering and algorithm selection strategies to model evaluation and design techniques specific to imbalanced datasets. We walk through a clinical example using readable code to highlight important considerations and common pitfalls in developing machine learning-based prediction models. We hope this practical guide facilitates developing and critically appraising robust clinical prediction models for the surgical community.

Serratus anterior plane block improves pain and incentive spirometry volumes in trauma patients with multiple rib fractures: a prospective cohort study.

Background: Rib fractures are common injuries associated with considerable morbidity, long-term disability, and mortality. Early, adequate analgesia is important to mitigate complications such as pneumonia and respiratory failure. Regional anesthesia has been proposed for rib fracture pain control due to its superior side effect profile compared with systemic analgesia. Our objective was to evaluate the effect of emergency physician-performed, ultrasound-guided serratus anterior plane block (SAPB) on pain and respiratory function in emergency department patients with multiple acute rib fractures. Methods: This was a prospective observational cohort study of adult patients at a level 1 trauma center who had two or more acute unilateral rib fractures. Eligible patients received a SAPB if an emergency physician trained in the procedure was available at the time of diagnosis. Primary outcomes were the absolute change in pain scores and percent change in expected incentive spirometry volumes from baseline to 3 hours after rib fracture diagnosis. Results: 38 patients met eligibility criteria, 15 received the SAPB and 23 did not. The SAPB group had a greater decrease in pain scores at 3 hours (-3.7 vs. -0.9; p=0.003) compared with the non-SAPB group. The SAPB group also had an 11% (CI 1.5% to 17%) increase in percent expected spirometry volumes at 3 hours which was significantly better than the non-SAPB group, which had a -3% (CI -9.1% to 2.7%) decrease (p=0.008). Conclusion: Patients with rib fractures who received SAPB as part of a multimodal pain control strategy had a greater improvement in pain and respiratory function compared with those who did not. Larger trials are indicated to assess the generalizability of these initial findings.

Thoracic and Lumbar Spine Injury: Evidence-Based Diagnosis, Management, and Outcomes.

BACKGROUND: Traumatic thoracolumbar spine injuries are associated with significant morbidity and mortality. Targeted for non-spine specialist trauma surgeons, this systematic scoping review aimed to examine literature for up-to-date evidence on presentation, management, and outcomes of thoracolumbar spine injuries in adult trauma patients. METHODS: This review was reported using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses checklist. We searched four bibliographic databases: PubMed, EMBASE, Web of Science, and the Cochrane Library. Eligible studies included experimental, observational, and evidence-synthesis articles evaluating patients with thoracic, lumbar, or thoracolumbar spine injury, published in English between January 1, 2010 and January 31, 2021. Studies which focused on animals, cadavers, cohorts with N <30, and pediatric cohorts (age <18 years old), as well as case studies, abstracts, and commentaries were excluded. RESULTS: A total of 2501 studies were screened, of which 326 unique studies were fully text reviewed and twelve aspects of injury management were identified and discussed: injury patterns, determination of injury status and imaging options, considerations in management, and patient quality of life. We found: (1) imaging is a necessary diagnostic tool, (2) no consensus exists for preferred injury characterization scoring systems, (3) operative management should be considered for unstable fractures, decompression, and deformity, and (4) certain patients experience significant burden following injury. DISCUSSION: In this systematic scoping review, we present the most up-to-date information regarding the management of traumatic thoracolumbar spine injuries. This allows non-specialist trauma surgeons to become more familiar with thoracolumbar spine injuries in trauma patients and provides a framework for their management.

Clinical prediction tool pitfalls and considerations: Data and algorithms.

In recent years, many surgical prediction models have been developed and published to augment surgeon decision-making, predict postoperative patient trajectories, and more. Collectively underlying all of these models is a wide variety of data sources and algorithms. Each data set and algorithm has its unique strengths, weaknesses, and type of prediction task for which it is best suited. The purpose of this piece is to highlight important characteristics of common data sources and algorithms used in surgical prediction model development so that future researchers interested in developing models of their own may be able to critically evaluate them and select the optimal ones for their study.

Development and Validation of a Model to Quantify Injury Severity in Real Time.

Importance: Quantifying injury severity is integral to trauma care benchmarking, decision-making, and research, yet the most prevalent metric to quantify injury severity-Injury Severity Score (ISS)- is impractical to use in real time. Objective: To develop and validate a practical model that uses a limited number of injury patterns to quantify injury severity in real time through 3 intuitive outcomes. Design, Setting, and Participants: In this cohort study for prediction model development and validation, training, development, and internal validation cohorts comprised 223 545, 74 514, and 74 514 admission encounters, respectively, of adults (age ≥18 years) with a primary diagnosis of traumatic injury hospitalized more than 2 days (2017-2018 National Inpatient Sample). The external validation cohort comprised 3855 adults admitted to a level I trauma center who met criteria for the 2 highest of the institution's 3 trauma activation levels. Main Outcomes and Measures: Three outcomes were hospital length of stay, probability of discharge disposition to a facility, and probability of inpatient mortality. The prediction performance metric for length of stay was mean absolute error. Prediction performance metrics for discharge disposition and inpatient mortality were average precision, precision, recall, specificity, F1 score, and area under the receiver operating characteristic curve (AUROC). Calibration was evaluated using calibration plots. Shapley addictive explanations analysis and bee swarm plots facilitated model explainability analysis. Results: The Length of Stay, Disposition, Mortality (LDM) Injury Index (the model) comprised a multitask deep learning model trained, developed, and internally validated on a data set of 372 573 traumatic injury encounters (mean [SD] age = 68.7 [19.3] years, 56.6% female). The model used 176 potential injuries to output 3 interpretable outcomes: the predicted hospital length of stay, probability of discharge to a facility, and probability of inpatient mortality. For the external validation set, the ISS predicted length of stay with mean absolute error was 4.16 (95% CI, 4.13-4.20) days. Compared with the ISS, the model had comparable external validation set discrimination performance (facility discharge AUROC: 0.67 [95% CI, 0.67-0.68] vs 0.65 [95% CI, 0.65-0.66]; recall: 0.59 [95% CI, 0.58-0.61] vs 0.59 [95% CI, 0.58-0.60]; specificity: 0.66 [95% CI, 0.66-0.66] vs 0.62 [95%CI, 0.60-0.63]; mortality AUROC: 0.83 [95% CI, 0.81-0.84] vs 0.82 [95% CI, 0.82-0.82]; recall: 0.74 [95% CI, 0.72-0.77] vs 0.75 [95% CI, 0.75-0.76]; specificity: 0.81 [95% CI, 0.81-0.81] vs 0.76 [95% CI, 0.75-0.77]). The model had excellent calibration for predicting facility discharge disposition, but overestimated inpatient mortality. Explainability analysis found the inputs influencing model predictions matched intuition. Conclusions and Relevance: In this cohort study using a limited number of injury patterns, the model quantified injury severity using 3 intuitive outcomes. Further study is required to evaluate the model at scale.

An Open-Source Curriculum to Teach Practical Academic Research Skills.

Academic productivity is important for career advancement, yet not all trainees have access to structured research programs. Without formal teaching, acquiring practical skills for research can be challenging. A comprehensive research course that teaches practical skills to translate ideas into publications could accelerate trainees' productivity and liberate faculty mentors' time. We share our experience designing and teaching "A Practical Introduction to Academic Research", a course that teaches practical skills including building productive habits, recognizing common statistical pitfalls, writing cover letters, succinct manuscripts, responding to reviewers, and delivering effective presentations. We share open-source educational material used during the Winter 2022 iteration to facilitate curriculum adoption at peer institutions.

FasterRib: A deep learning algorithm to automate identification and characterization of rib fractures on chest computed tomography scans.

OBJECTIVE: Characterizing and enumerating rib fractures are critical to informing clinical decisions, yet in-depth characterization is rarely performed because of the manual burden of annotating these injuries on computed tomography (CT) scans. We hypothesized that our deep learning model, FasterRib , could predict the location and percentage displacement of rib fractures using chest CT scans. METHODS: The development and internal validation cohort comprised more than 4,700 annotated rib fractures from 500 chest CT scans within the public RibFrac. We trained a convolutional neural network to predict bounding boxes around each fracture per CT slice. Adapting an existing rib segmentation model, FasterRib outputs the three-dimensional locations of each fracture (rib number and laterality). A deterministic formula analyzed cortical contact between bone segments to compute percentage displacements. We externally validated our model on our institution's data set. RESULTS: FasterRib predicted precise rib fracture locations with 0.95 sensitivity, 0.90 precision, 0.92 f1 score, with an average of 1.3 false-positive fractures per scan. On external validation, FasterRib achieved 0.97 sensitivity, 0.96 precision, and 0.97 f1 score, and 2.24 false-positive fractures per scan. Our publicly available algorithm automatically outputs the location and percent displacement of each predicted rib fracture for multiple input CT scans. CONCLUSION: We built a deep learning algorithm that automates rib fracture detection and characterization using chest CT scans. FasterRib achieved the highest recall and the second highest precision among known algorithms in literature. Our open source code could facilitate FasterRib's adaptation for similar computer vision tasks and further improvements via large-scale external validation. LEVEL OF EVIDENCE: Diagnostic Tests/Criteria; Level III.