The pruning objective has recently extended beyond accuracy and sparsity to robustness in language models. Despite this, existing methods struggle to enhance robustness against adversarial attacks when continually increasing model sparsity and require a retraining process. As humans step into the era of large language models, these issues become increasingly prominent. This paper proposes that the robustness of language models is proportional to the extent of pre-trained knowledge they encompass. Accordingly, we introduce a post-training pruning strategy designed to faithfully replicate the embedding space and feature space of dense language models, aiming to conserve more pre-trained knowledge during the pruning process. In this setup, each layer's reconstruction error not only originates from itself but also includes cumulative error from preceding layers, followed by an adaptive rectification. Compared to other state-of-art baselines, our approach demonstrates a superior balance between accuracy, sparsity, robustness, and pruning cost with BERT on datasets SST2, IMDB, and AGNews, marking a significant stride towards robust pruning in language models.

Towards Robust Pruning: An Adaptive Knowledge-Retention Pruning Strategy for Language Models | VIDEO

We address the task of identifying euphemistic abuse (e.g. "You inspire me to fall asleep") paraphrasing simple explicitly abusive utterances (e.g. "You are boring"). For this task, we introduce a novel dataset that has been created via crowdsourcing. Special attention has been paid to the generation of appropriate negative (non-abusive) data. We report on classification experiments showing that classifiers trained on previous datasets are less capable of detecting such abuse. Best automatic results are obtained by a classifier that augments training data from our new dataset with automatically-generated GPT-3 completions. We also present a classifier that combines a few manually extracted features that exemplify the major linguistic phenomena constituting euphemistic abuse.

Euphemistic Abuse – A New Dataset and Classification Experiments for Implicitly Abusive Language | VIDEO

Trained medical interpreters are crucial to ensure that patients who speak English with limited proficiency (LEP) receive high-quality care
Current medical education lacks formal training that incorporates practice in simulated environments 
We sought to evaluate MS1 and MS2 ability to use interpreters and address knowledge gaps through near-peer training sessions

Analyzing Medical Students’ Preparation and Confidence with Interpreters

**Scoring is only possible in the semifinals hall**

**Abstract:**



Uncovering Patterns between Obstructive Sleep Apnea and Disease Comorbidities: An Exploration of Patient Demographics

Background: Estimated Glomerular Filtration Rate (eGFR) is the standard method for determining kidney function and chronic kidney disease (CKD) stage classification. In the Spina Bifida (SB) population, eGFR estimates from a creatinine-based equation may not accurately represent kidney function due to the lower muscle mass in this population. We hypothesized that estimating equations utilizing Cystatin-C (CysC) would yield different CKD classifications than Creatinine (Cr) – based equations for the adult SB population. 

Methods: We conducted a retrospective chart review of all patients seen in an adult urology clinic with a diagnosis of spina bifida from 2004-2021 that were managed in our urology clinic with creatinine and cystatin C values available. All patients were > 18 years old. A cross-sectional study of patients with a CysC and concurrent Cr measurement was performed; the most recent lab values were utilized if multiple existed. Three adult eGFR equations were utilized. Patients were staged based on CKD classification (GFR > 90, GFR 60-90, GFR <60). A Chi-square test was conducted to determine the difference in CKD classification based on each equation.

Results: A total of 105 SB patients (63% female) were included in cross-sectional analysis. The mean age was 33 ± 6.5 years. Medical comorbidities including diabetes (12.4%), hypertension (28.6%), obesity (34.2%) were noted. Table 1 shows the difference in eGFR and CKD classifications between equations. On Chi-square analysis, CKD classifications were significantly different by the type of equation used (p = 0.007). Pair-wise comparison showed that significance was driven by the difference between cystatin C and creatinine equations (p = 0.002) and the cystatin C and combined equations (p = 0.05). 

Conclusion: In an adult SB population, CysC-based equations result in lower eGFRs and different CKD classifications. Determining an appropriate eGFR measurement is essential to assessing and managing renal function in these medically complex patients.

Variability in Cystatin C and Creatinine-Based equations for Chronic Kidney Disease Staging in Adult Spina Bifida Patients

Background
Patients who present with cervical dystonia may also present with cervical myelopathy. This presentation can be problematic because patients with cervical myelopathy need fusion surgery such as an anterior discectomy and fusion (ACDF) to stabilize the spine and prevent damage to the spinal cord. The involuntary tremor caused by cervical dystonia can result in problems with boney fusion in the post-operative period. Deep Brain Stimulation (DBS) provides a potential solution to this problem when more conservative treatments like medication, halo vests, and botulinum toxin fail to reduce tremor.
Case Presentation
A 46-year-old female presented to our center with a chief complaint of neck pain. The patient also presented with bilateral upper extremity radiculopathy, paresthesia, and a myelopathic gait. She had a 15-year history of cervical dystonia and had tried various medications and botulinum toxin injections without resolution of symptoms. In addition to this the patient had cervical myelopathy caused by disc herniation at the C5-C6 and C6-C7 level. Although the patient was a candidate for ACDF, she was advised to have her dystonic tremor treated before ACDF to prevent pseudoarthrosis and hardware failure because of the tremor. The patient underwent bilateral DBS in the Globus pallidus interna (GPi) utilizing an Abbott InfinityTM DBS system. After placement of the DBS-GPi leads, batteries were placed, and the system was turned on which was able to provide significant relief of the patient’s tremor. The patient was followed up at 3 months and had a >75% reduction in her tremors. After the significant reduction in tremor, the patient underwent C5-C6 and C6-C7 ACDF at 6 months post DBS-GPi. The patient reported that the paresthesia had improved significantly, and the weakness had gradually improved at 3 months post ACDF. There was no radiographic evidence of hardware loosening or malalignment.
Discussion
The presentation of both Cervical Dystonia and Cervical Myopathy concurrently is a challenging couple of pathologies to manage. Bilateral DBS-GPi prior to ACDF is a viable strategy for managing these pathologies surgically. 


DBS before ACDF for a patient with Cervical Dystonia and Cervical Myelopathy

Background: Testicular infarction is a rare complication of epididymitis. We present a case of testicular infarction to highlight this complication and the role of ultrasound in aiding diagnosis to guide management.

Case Presentation: 26-year-old male, with recent admission for left epididymitis secondary to Multiple Drug Resistant Enterobacter, presented to the emergency department with a two-day history of progressively worsening left scrotal pain and swelling. Physical examination revealed left scrotal erythema and tenderness. Scrotal ultrasound demonstrated asymmetric enlargement and heterogeneous appearance of the left testicle consistent with left scrotal edema. The left testicle showed peripheral hypervascularity without intrinsic flow on color and spectral Doppler interrogation compatible with testicular infarction. In contrast, the right testicle was homogenous and with normal arterial and venous spectral Doppler flow. Management included intravenous antibiotic and urology consultation with subsequent partial orchiectomy. Patient’s hospital course was complicated by wound dehiscence and residual necrotic tissue requiring scrotal debridement and washout. After transition to oral antibiotics, he was discharged with outpatient follow-up with urology.

Discussion: Ultrasound with color flow Doppler is the primary method to diagnose testicular infarction secondary to epididymitis. Characteristics of testicular infarction include asymmetrically decreased or absent intrinsic testicular blood flow with or without increased peripheral flow. Although understanding of the pathophysiology of testicular infarction is evolving, arterial and venous insufficiency are suspected. Risk of global testicular infarction is low due to the testes receiving blood supply from the testicular, the deferential and cremasteric arteries. Persistent clinical symptoms and worsening imaging findings can warrant surgical intervention. In addition to spectral and color doppler, there is expanding utilization of contrast- enhanced sonography (CEUS) to increase sensitivity for testicular infarction. Testicular infarction is a rare but important diagnosis to consider in the setting of epididymitis that has management implications to prevent morbidity.

Treatment Failure of Epididymitis leads to Testicular Infarction

Purpose: The adoption of artificial intelligence (AI) in image generation raises concerns regarding potential bias, as these technologies may not be reflective of the underlying population in the images they generate. Our study investigates whether AI-generated images of medical students accurately represent the diversity of the current student population.

Methods: Using the DALL-E (Open AI) image generation algorithm, we created 300 images with the text prompt "medical student." These images were independently analyzed by two researchers for demographic indicators including sex, race/ethnicity, age group, setting, and attire. We summarized the results with descriptive statistics and subgroup analyses were performed to analyze differences in portrayals based on student sex and race/ethnicity. The demographic proportions in the virtual cohort of images was graphically compared with AAMC medical school enrollment data of actual medical students.

Results: Of the 300 generated images, 227 (76%) were females and 223 (74%) were White, indicating over-representation compared to actual medical school demographics. Images of Black and Latino/Hispanic students were more commonly depicted in scrubs compared to White students, who were often portrayed in white coats or collared shirts (p=0.002). No representation was found for Native American/Alaskan Native or Native Hawaiian/Pacific Islander students.

Conclusion: Our study reveals significant disparities in the representation of demographic groups in AI-generated images of medical students, indicating potential bias in these technologies. These biased portrayals could perpetuate existing stereotypes and hinder diversity efforts. Future research should aim to further identify and address these biases to promote more equitable applications of AI tools.

Portrayal of Medical Students in Artificial Intelligence Generated Images

Introduction:
High-fidelity three-dimensional (3D) models provide a better understanding of spatial relationships among complex anatomical structures. Medical students often lack hands-on opportunities to develop an in-depth understanding of applied anatomy prior to their clinical years, a trend exacerbated by the COVID-19 pandemic. There is limited research on the utilization of 3D models in the clinical years. This study evaluates the effectiveness of an interactive 3D-printed inguinal hernia model in teaching applied surgical anatomy to third-year medical students (MS3).

Materials and Methods:
3D-printed models of inguinal anatomy were created using de-identified CT scans, 3D-modeling software, and 3D-printing services. A single-blind randomized control trial of third-year medical students was conducted at a single academic institution. The tactile-group received three educational tools describing inguinal hernia anatomy and repair: an established video module (WISE-MD), an illustrated procedure guide, and a 3D-printed model. The non-tactile group was provided the WISE-MD module and the illustrated procedure guide. Both groups received pre- and post-tests on basic and applied anatomy. Statistical analysis was performed using a t-test or Mann–Whitney U test for continuous variables and Chi-square or Fisher's exact tests for categorical variables. Statistical significance was defined by p <0.05.

Results:
One hundred and three third-year medical students participated in this study. There were no differences in pre-test scores between study groups (p= 0.98). The tactile-group scored better on post-test questions on applied anatomy (p= 0.03), but no differences were noted in scores on basic anatomy questions (p= 0.41). When comparing performance between overall pre-test and post-test scores, the tactile group demonstrated improvement in test scores which trended toward significance (p= 0.09).

Conclusions:
Third-year medical students using high-fidelity 3D-printed models demonstrated improved test scores on applied anatomy when compared to students who were exposed to conventional educational resources. With the rising use of virtual learning environments, interactive learning tools like 3D-printed anatomy models may have an important role in knowledge acquisition during the surgical clerkship.

3D Printing in Surgical Education:A Visuospatial and Kinesthetic Learning Module of Inguinal Hernia Repair for Third-Year Medical Students

Objective: Salivary gland tumors are a heterogenous group of lesions with variable pathology and clinical outcomes. Most published data is derived from studies conducted at tertiary care centers. Our study analyzed the experience from a community setting to determine significant differences, if any, in pathological distribution and clinical outcomes compared to the existing literature.

Methods: We performed a retrospective analysis of all salivary gland tumors that presented to a large community practice over a 20-year period. Retrospective chart analysis was performed for demographics, clinical presentation, imaging, cytology, histopathology, and clinical outcome data.

Results: Of 806 patients, the parotid gland was the most common site in 683 patients (84.7%), followed by submandibular in 78 (9.7%) and sublingual in 45 (5.6%). 203 patients were managed conservatively without definitive diagnosis or lost to follow up. A total of 495 patients underwent surgical intervention within the community practice. 26 patients underwent surgical excision at an outside hospital. 82 patients were determined to have a benign diagnosis based on FNA or excisional biopsy alone. Final histopathology was benign in 505 cases (83.7%), while 98 tumors (16.3%) received a diagnosis of primary or secondary malignancy. For the parotid gland, pleomorphic adenoma (155) and Warthin’s tumor (155) were the most common benign diagnoses, while mucoepidermoid carcinoma (13), adenocarcinoma (8), and acinic cell carcinoma (8) were the most common primary malignancies. 

Conclusions: We found a higher rate of benign tumor pathology compared to the existing literature. While the outcome data on surgical treatment of benign tumors is comparable to the existing literature, the same conclusion cannot be drawn for malignant tumors, given relatively small numbers in our series and likely disparity in the complexity of the surgical cases in tertiary care centers.

Salivary Gland Tumors: A Twenty-Year Review from a Single Community Practice

Introduction
The external urethral sphincter reflex is a new neurmonitoring modality used to protect bowel, bladder, and sexual function during surgery. EUSR and bulbocavernosus reflex (BCR) assess the sacral segments of the spinal cord, cauda equina, and sacral nerve roots and provides a functional evaluation of the S2-S4 mediated reflex arc.
Objectives
We present a case study where we demonstrate for the first time the EUSR prognosticating postoperative bladder function independently of bowel function.
Methods
A 59-year-old female with hypertension and type 2 diabetes mellitus presented with a tumor causing cord compression at the T10 level. She underwent a T9-T11 laminectomy for tumor resection with a multimodal neuromonitoring paradigm including motor evoked potentials, somatosensory evoked potentials, EMG, BCR, and EUSR. The BCR and EUSR were activated using needle electrodes placed in the clitoris (cathode) and labia majora (anode) to activate the afferent sensory fibers of the dorsal clitoral nerves. Needle electrodes were placed in the left and right external anal hemisphincters to record the downstream motor component of the BCR. A specialized foley catheter embedded with 3D printed electrodes was used to record from the external urethral sphincter muscle for the EUSR.
Results
Baseline neuromonitoring data was acquired in all modalities except the EUSR. The
laminectomy and tumor resection were uneventful, and almost immediately following tumor removal, the EUSR began to reproduce and remained stable throughout the remainder of the procedure. The final pathology was benign schwannoma, and the patient was discharged intact on post-operative day 8. Follow-up at one month revealed cured urinary incontinence.
Conclusion
While this case study has limitations, it suggests that EUSR may not only provide redundancy to BCR monitoring but may discretely assess pathways associated with bladder function and prognosticate post-operative outcomes. More research with a larger cohort is needed to prove efficacy and statistical significance.

Prognostic Value of Intraoperative External Urethral Sphincter Reflexes in
Intradural Spinal Cord Surgery: A Case Report

Background
Type 2 diabetes mellitus (T2DM) is a global epidemic affecting approximately 9.3% of the world population and has an all-cause mortality rate of 8.5%. An increasing number of patients with chronic conditions access online resources for educational purposes, with recent statistics suggesting that up to 80% of adults in the United States have obtained medical information through online search engines. This study aims to assess the accuracy of the information offered by artificial intelligence (AI) chatbots to T2DM questions using ChatGPT.

Methods:
An initial set of questions about T2DM and its complications were obtained from an online platform such as American Diabetes Association. Subsequently, these questions underwent a thorough evaluation by physicians well-versed in diabetes management before being presented to ChatGPT. Each question was introduced to ChatGPT three times, and the responses were recorded. Two study staff reviewed the three sets of responses and summarized the majority response that was repeated at least twice. Data were reviewed by two board-certified internal medicine physicians with a minimum of 11 years of experience. Responses were categorized into appropriate; provide accurate information that meets the standard of care, inappropriate; contain erroneous information but meets the minimal standard of care, and unreliable; the response provides inconsistent content that does not meet the standard of care.


Results:
Of the original 110 questions, 70 were included in the final dataset. Out of the 70 summarized responses generated by ChatGPT, 89.3% were categorized as appropriate, indicating that the AI model provided accurate information meeting the standard of care for Type 2 diabetes management and its associated complications. On the other hand, 10% of the responses were deemed inappropriate, containing errors but still meeting the minimal standard of care in addressing the questions. The remaining 0.7% of ChatGPT's responses were classified as unreliable, displaying inconsistent content that did not meet the standard of care in terms of accuracy and reliability


Conclusion
ChatGPT offers a promising source of information for educating patients with T2DM. The study demonstrates high accuracy in providing appropriate responses that align with the standard of care, positioning it as a useful tool in enhancing patient education and diabetes management.


Underline digital video library

Towards Robust Pruning: An Adaptive Knowledge-Retention Pruning Strategy for Language Models | VIDEO

Euphemistic Abuse – A New Dataset and Classification Experiments for Implicitly Abusive Language | VIDEO

Analyzing Medical Students’ Preparation and Confidence with Interpreters

Uncovering Patterns between Obstructive Sleep Apnea and Disease Comorbidities: An Exploration of Patient Demographics

Variability in Cystatin C and Creatinine-Based equations for Chronic Kidney Disease Staging in Adult Spina Bifida Patients

DBS before ACDF for a patient with Cervical Dystonia and Cervical Myelopathy

Treatment Failure of Epididymitis leads to Testicular Infarction

Portrayal of Medical Students in Artificial Intelligence Generated Images

3D Printing in Surgical Education:A Visuospatial and Kinesthetic Learning Module of Inguinal Hernia Repair for Third-Year Medical Students

Salivary Gland Tumors: A Twenty-Year Review from a Single Community Practice

Prognostic Value of Intraoperative External Urethral Sphincter Reflexes in Intradural Spinal Cord Surgery: A Case Report

Artificial Intelligence Shifting the Paradigm of Diabetes Care

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES