United States

Background
Chat GPT is a natural language processing model that has been evaluated and used in a variety of different applications. However, its ability to provide references for the appropriate use of immunohistochemistry staining in dermatopathology has not been evaluated.

Methods
Both ChatGPT v3.5 and v4.0 were evaluated. Data collection occurred from May through June 2023. A total of 33 diagnoses involving benign growths, carcinomas, benign and malignant adnexal tumors, and cutaneous lymphomas were queried. For each condition or differential, ChatGPT was queried with “I am a dermatopathologist trying to work up a skin biopsy case with immunohistochemistry. I suspect the diagnosis to be _. Please supply a list of references that I can use to evaluate the utility of specific immunohistochemical stains for this diagnosis.” For ChatGPT v3.5, each prompt was regenerated 10 times; for ChatGPT v4.0, each prompt was regenerated 5 times. The outputs were recorded and compiled.

Results
ChatGPT v3.5 rarely provided clinically useful citations (15.0%), and over half were confabulated (51.5%). In ChatGPT v4.0, performance was improved with a greater proportion of clinical relevant, useful citations (23.5%) and much less confabulated references (27.0%). However, there was still a substantial portion of irrelevant references (22.9%), general textbooks (16.2%), and instances where ChatGPT failed to even provide references (9.0%).

Conclusion
ChatGPT can provide clinically useful references, but significant supervision and validation is required for the tendencies of ChatGPT to fabricate material. Future natural language processing models designed to augment diagnostic medicine may improve dermatopathologists&#39; ability to find clinically useful references.

2023 AMA Research Challenge

Evaluating the diagnostic utility of ChatGPT for providing dermatopathology immunohistochemistry reference recommendations

Background
Chat GPT is a natural language processing model that has been evaluated and used in a variety of different applications. However, its ability to provide references for the appropriate use of immunohistochemistry staining in dermatopathology has not been evaluated.

Methods
Both ChatGPT v3.5 and v4.0 were evaluated. Data collection occurred from May through June 2023. A total of 33 diagnoses involving benign growths, carcinomas, benign and malignant adnexal tumors, and cutaneous lymphomas were queried. For each condition or differential, ChatGPT was queried with “I am a dermatopathologist trying to work up a skin biopsy case with immunohistochemistry. I suspect the diagnosis to be _. Please supply a list of references that I can use to evaluate the utility of specific immunohistochemical stains for this diagnosis.” For ChatGPT v3.5, each prompt was regenerated 10 times; for ChatGPT v4.0, each prompt was regenerated 5 times. The outputs were recorded and compiled.

Results
ChatGPT v3.5 rarely provided clinically useful citations (15.0%), and over half were confabulated (51.5%). In ChatGPT v4.0, performance was improved with a greater proportion of clinical relevant, useful citations (23.5%) and much less confabulated references (27.0%). However, there was still a substantial portion of irrelevant references (22.9%), general textbooks (16.2%), and instances where ChatGPT failed to even provide references (9.0%).

Conclusion
ChatGPT can provide clinically useful references, but significant supervision and validation is required for the tendencies of ChatGPT to fabricate material. Future natural language processing models designed to augment diagnostic medicine may improve dermatopathologists' ability to find clinically useful references.

poster

Welcome to the **largest national, multi-specialty research** event featuring research from nearly 700 medical students, residents, fellows, and international medical graduates.

Use the buttons below or navigation menu to access the event.
 


[![](https://assets.underline.io/uploads/markdown_image/1/image/3e33a8c754cc49610b30950617d813e7.png)](https://underline.io/events/434/sessions?searchGroup=all)
 
[![](https://assets.underline.io/uploads/markdown_image/1/image/551404d429c4e274f9a54cbdd7081918.png)](https://underline.io/events/434/posters?searchGroup=all)
 
[![](https://assets.underline.io/uploads/markdown_image/1/image/677d92f566b880cebd36905fc738fc10.png)](https://underline.io/events/434/lounge)
 
[![](https://assets.underline.io/uploads/markdown_image/1/image/c7392833c7261c3874c5b43fa2f4df19.png)](https://underline.io/events/434/help-desk)
 

#### Event Details

The 2023 American Medical Association Research Challenge includes a poster symposium hall and a semifinals hall with the top scored research. **Log in at any time throughout the event to:**

* **View posters and presentations** on a variety of topics and specialties
* **Score posters in the semifinals** to help decide the top five finalists (AMA member exclusive opportunity)
* **Connect with participants** and provide feedback to help advance their research
* **Attend education sessions** with physician experts on key research topics

At the conclusion of this event, the AMA will announce the top five finalists who will advance to the AMA Research Challenge finals, present their research to an elite panel of judges, and compete to win a $10,000 grand prize sponsored by Laurel Road. 

### Grand Prize Sponsor

AMA would like to thank Laurel Road for sponsoring a $10,000 grand prize for the winner of the AMA Research Challenge. Laurel Road is not involved in the judging or scoring process to select the winner. The winner will be announced on February 4, 2024 during the AMA Research Challenge finals. Payment of the grand prize will be directly from Laurel Road. See the AMA Research Challenge Official Rules for terms and conditions.

![](https://assets.underline.io/uploads/markdown_image/1/image/611e331d4a0eba0920f240cfc04b4d77.png)

**There is no registration fee for this event! **

Register to explore research in a variety of topics. AMA members get exclusive access to score posters in the semifinals.

**Please use this link [here](https://www.ama-assn.org/about/events/ama-research-challenge) to register for free.**

Registration Is Required

Join the AMA Research Challenge poster symposium and semifinals to explore top research posters, score posters in the semifinals (an AMA member exclusive opportunity) and attend education sessions with experienced physicians to learn tips to advance your research and career. 

Background

Tumor-to-tumor metastasis (TTM) is an exceptionally rare phenomenon, especially so when involving cutaneous neoplasms. Therefore, they may present a diagnostic conundrum for pathologists.

TTM are defined by the following criteria:

1- More than one primary tumor must exist.

2- The host tumor is a true neoplasm.

3- The primary tumor shows growth inside the secondary tumor, not just transitional growth indicative of collision.

4- The host tumor cannot be a lymph node affected by leukemia or lymphoma.

This systematic review aims to characterize the cutaneous TTM as both donor and recipient tumors. This will provide information for pathologists to consider in the diagnosis of TTM. Additionally, these events may provide insight into underlying mechanisms of tumor tropism in metastatic disease.

Methods

PubMed, Embase, Medline, and ScienceDirect were searched using the terms “tumor-to-tumor metastasis” and “cutaneous””. Titles and abstracts were reviewed independently by 2 reviewers (NC, MRM). Inclusion criteria required English language and peer-reviewed studies. Consensus was reached for all articles. Furthermore, citations of all included studies were

reviewed to find further cases. The following data were extracted: title, patient age, sex, race, donor and recipient tumor type, size, location, subtype, and special features, treatment history, and metastatic burden.

Results

13/673 articles met inclusion criteria and were further reviewed. The most common donor cutaneous malignancy was melanoma (7/13), The most common recipient tumor of cutaneous metastatic melanoma was basal cell carcinoma (2/13). Lipomas and basal cell carcinomas(BCCs) were the two most common recipient cutaneous tumors (2/13 each), and the phenomena often occurred in the setting of widespread metastasis (7/13 cases).

Conclusion

While uncommon, tumor-to-tumor metastasis should be considered as an important differential diagnosis that may have profound implications for treatment and prognosis. Melanomas were implicated in cutaneous TTM more often than any other tumor. This information will be useful for pathologists to consider in working up challenging cases suggestive of tumor-to-tumor metastasis.

Characteristics of TTM identified in this study may elucidate mechanisms underlying metastatic tropism. It is unclear to what degree TTM occurs from chance, or how the host tumor microenvironmental milieu affects the probability of metastasis. Our findings suggest host tumor microenvironment plays an important role in TTM, as only 7/13 of cases occurred in patients with diffuse metastatic burdens. Additional studies teasing apart the seed-and-soil hypothesis of TTM are needed to understand this phenomenon use it to improve diagnosis, prognosis, and treatment of patients with metastatic disease.

Cutaneous Neoplasms as Donors and Recipients in Tumor-to-Tumor Metastasis: A Systematic Review of the Literature

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

**Abstract:**



Advancing Cancer Diagnostics with Artificial Intelligence (AI): A Comprehensive Systematic Review

Elevated Infection Risk in Post-Transplant 
Lymphoproliferative Disorder: the Role for Prophylaxis


Background
For patients undergoing radical prostatectomy for prostate cancer (PCa), accurate risk stratification is essential to guide post-prostatectomy decision making. Recently, there has been success in using multi-modal artificial intelligence (AI) to aid in risk stratification. Herein, we present a TRansfer learning-based multi-modal Artificial InteLligence (TRAIL) model for post-prostatectomy biochemical recurrence (BCR) risk stratification using pre-prostatectomy data only.

Methods
Patients contained within a prospective localized PCa registry at a single institution were utilized. Digital pathology slides from the diagnostic biopsies prior to radical prostatectomy were scanned at 20x resolution. Features were extracted from pathology slides via 2 transfer learning steps: (1) Inception-v3 network trained on the PCa PANDA dataset to generate a heatmap corresponding to the tumor areas followed by (2) AlexNet network that learns 256 AI representative features from each slide. Afterwards, TRAIL combines AI + clinical features using random under-sampling and the standard boosting procedure AdaBoost (RUSBoost), and train a classification ensemble model based on weak tree learners to predicts 2yr-BCR. TRAIL training was performed on 245 patients and was then locked and applied to the test set of 119 patients. Accuracy and the area under the curve (AUC) were calculated and compared to CAPRA-S.
After training, the occlusion sensitivity algorithm was used to localize the areas within the slide driving the AI features. Lastly, the visual features areas were segmented using K-mean clustering with 4 classes (background, basal cells, acinar cells, and fibromuscular tissue), and interpretable qualitative features were extracted. 

Results
A total of 808 digital whole pathology biopsy slides from 364 patients treated with radical prostatectomy were included. The median follow-up was 48 months (Range: 1-132 months). The cumulative incidence of 2-and 5-year BCR were 11% and 18% respectively. Clinical factors included age, PSA, and fraction of positive biopsy cores. On the testing set, TRAIL achieved a 2-year BCR AUC of 0.81 and accuracy of 0.95, and was superior to CAPRA-S (AUC = 0.57). 
Qualitative studies revealed nests and cords of cells with vague lamina formation and scattered clear vacuoles. Further histopathological review is undergoing.

Conclusion
Through a combination of deep and ensemble learning, TRAIL incorporates clinical and histopathology features, enabling an improved BCR risk stratification post-prostatectomy. In contrast to commonly used biomarkers, TRAIL is independent of prostatectomy data, and offers patients the advantage of decision-forming about prostatectomy prior to undergoing surgery. Future work with larger datasets is warranted to further optimize TRAIL for clinical use.

TRAIL: A Diagnostic Biopsy-derived, Artificial Intelligence-based, End-to-end, Automated, Interpretable, and Predictive Biomarker for Biochemical Recurrence after Prostatectomy

Background
Prostate cancer (PCa) risk is understood to be mostly unmodifiable and inherited, but there is evidence that environmental and behavioral factors may also contribute. A recent study of health professional cohorts suggests a healthy lifestyle can mitigate a high inherited risk of lethal PCa. It is unknown how modifiable factors affect PCa risk in more diverse populations. The objective of this study was to determine the effects of healthy lifestyle on PCa risk when accounting for race/ethnicity, family history, and genetic risk in a diverse population.

Methods
The Million Veteran Program (MVP) is a national, population-based cohort study of United States military veterans conducted 2011-2021 with 590,750 male participants available for analysis. A healthy lifestyle score (range 0-3) was calculated with a point assigned for each of the following at MVP enrollment: not a current smoker, body mass index (BMI) ≤30 and strenuous activity ≥2 days per week. Genetic risk was assessed via a previously validated polygenic hazard score using genotype data from 290 variants. Age at diagnosis of PCa, diagnosis of metastatic PCa, and death from PCa were assessed via Cox proportional hazards models.

Results
Healthy lifestyle score (HLS) was associated with lower risk of metastatic PCa (HR per unit/point of HLS: 0.77, 95% CI 0.73–0.81, p<0.001) and of fatal PCa (HR 0.73, 95% CI 0.67–0.81, p<0.001) in univariable analysis. Black participants with higher HLS similarly had reduced risk of metastatic and fatal PCa. Analysis of HLS components showed smoking to have a large effect on PCa risk: non-smokers were less likely to be diagnosed with PCa (OR 0.89, 95% CI 0.86–0.92, p<0.001) and far less likely to develop metastatic PCa (OR 0.50, 95% CI 0.46–0.55, p<0.001) or fatal PCa (OR 0.33, 95% CI 0.28–0.39, p<0.001). When accounting for race/ethnicity and family history in multivariable analysis, HLS was independently associated with reduced metastatic (HR 0.82, 95% CI 0.77–0.87, p<0.001) and fatal PCa (HR 0.76, 95% CI 0.68–0.86, p<0.01) but not PCa diagnosis (HR 1.01, 95% CI 0.99–1.03). When accounting for genetic risk (PHS290) as well as race/ethnicity and family history, HLS remained independently associated with reduced metastatic and fatal PCa. 

Conclusion
Adherence to a healthy lifestyle is associated with reduced risk of metastatic or fatal PCa, which offsets inherited risk.

A Healthy Lifestyle Modifies Prostate Cancer Risk in the Million Veteran Program

Background: Depression is a major contributor to disability worldwide, and adults frequently present to primary care clinicians for behavioral health treatment. In previous studies, the Collaborative Care Model (CoCM) significantly improved participants’ depressive symptoms, decreased time to symptom remission, and lowered risk for CVD, all while remaining cost-effective. To successfully translate this program into practice, an understanding of the determinants of program implementation, particularly the roles and characteristics of the individuals involved, is essential. 

Methods: This study was a secondary analysis of the Collaborative Behavioral Health Program (CBHP) study (NCT04321876) in which CoCM was implemented in 11 primary care clinics. Interviews were conducted 4- and 15-months after implementation with behavioral care managers, practice managers, and primary care clinicians. We used Framework-Guided Rapid Analysis with the Consolidated Framework for Implementation Research, Version 2.0 (CFIR 2.0; Individuals domain), to analyze response data. 

Results: Mid-level leaders (practice managers) were enthusiastic about CBHP (Motivation), appreciated integration of behavioral health services into primary care (Need), had time to assist clinicians (Opportunity), and relayed team member communication (Capability). Although CBHP reportedly lessened burden for implementation team members (clinicians, staff; Need), some implementation team members were hesitant to reallocate patient care (Motivation). Innovation deliverers (behavioral care managers) were eager to deliver CBHP (Motivation) and confident in assisting patients (Capability); their opportunity to deliver CBHP could be limited by clinician referrals (Opportunity). CBHP alleviated access barriers faced by innovation recipients (primary care patients; Need) but it was difficult to secure services for severe disorders (Capability) and those with less insurance coverage (Opportunity).      

Conclusion: Understanding the individual roles and characteristics that influence implementation is a key domain for implementation engagement and adoption. The CFIR 2.0 Individuals domain provided a deep and nuanced understanding of the ways in which the needs, capabilities, opportunities, and motivations of multi-level actors determined the success of CoCM implementation. Overall, participants were in favor of CoCM and highlighted the positive impacts of CBHP from the perspective of the clinic, health care team, and patients. Participants emphasized the benefits of integrating behavioral health services into primary care, and the ways in which CBHP lessened the burden on clinicians while providing patients with comprehensive care. Future CoCM implementation should include strategies focused on education and training, encouraging clinician buy-in, and preparing referral paths for patients with severe or diverse disorders.

Multilevel Perspectives of Implementing Integrated Behavioral Health in Primary Care

Introduction:
In the context of patient safety in hospitals, a significant number of adverse events are identified, with a considerable portion being preventable. Timely detection of potential clinical deterioration events (CDEs) is crucial for mitigating or preventing adverse outcomes. The availability of electronic health records (EHRs) has facilitated the implementation of Artificial Intelligence (AI) algorithms known as "early warning systems" (EWS) for early detection. For an EWS to effectively improve outcomes, it must accurately predict adverse outcomes, do so ahead of clinicians and provide sufficient lead time for intervention. In this study, we evaluated the performance of a widely used EWS, the Epic Deterioration Index (DI), in terms of its ability to provide early warnings before clinicians and its lead time for intervention.

Methods:
Our study was conducted at a large academic medical center. The DI operated in the background, without direct visibility to frontline clinicians. This allowed us to compare the performance of unaided clinicians with the DI concurrently. We analyzed inpatient admissions from 03/01/2022 to 08/31/2022. Out of 21,880 admissions, 18,431 were excluded based on non-CDE hospital course, and 2,813 were excluded as ICU transfers. From the remaining 636 patients, we selected those who experienced a rapid response call or a cardiopulmonary arrest and determined the occurrence of definite or likely CDEs based on established criteria. We retrospectively examined the 72-hour period preceding the CDE to identify the time when clinicians initially recognized the potential clinical deterioration through order and documentation reviews. The lead time for unaided clinicians was calculated based on the first instance of recognition. The DI, running invisibly in the background, assigned patients to low, moderate, or high-risk categories for experiencing a CDE. We identified the first transition from low to moderate/high risk that persisted for at least 1 hour as a signal for potential CDE and used it to calculate the lead time provided by the DI. Comparative analyses were conducted using chi-square and independent t-test calculations.

Results:
Out of the 100 patients who experienced definite or likely CDEs, 81 cases were recognized by either clinicians or DI. The DI identified a CDE in 37 cases, while unaided clinicians detected potential CDEs in 57 cases (χ2 = 8.0, p < 0.01). On average, clinicians had a lead time of 9.13 hours, while the DI provided a lead time of 4.53 hours (p < 0.01). Thus, clinicians detected potential CDEs, on average, 4.53 hours earlier (95% CI [1.44, 7.76]).

Conclusion:
Our findings indicate that clinical judgment has not been entirely superseded by commonly used AI. Unaided clinicians detected potential CDEs more frequently and had significantly longer lead times compared to the DI. However, a notable minority of cases demonstrated that the DI outperformed clinicians in terms of speed. Therefore, while AI should not replace human efforts, it can be a valuable adjunct if appropriately implemented.

The Race to Save Lives: A Comparative Analysis of Lead Time for Detecting Clinical Deterioration between Clinicians and Artificial Intelligence

Background
Dietary recall assessments are important clinical tools used by Registered Dietitians (RDs) for assessing eating habits, malnutrition and obesity risk, and health status. Current methods of dietary recall assessment pose barriers to accurately assessing the nutritional intake of older adults due to age-related increases in risk for cognitive decline and more complex health histories. Voice-assistant systems (VAS) are emerging technologies that have been shown to improve the health of older adults. Our study explored the challenges older adults face with dietary recall and whether VAS integration could improve current dietary recall methods from the perspective of RDs.

Methods
This qualitative study utilized semi-structured interviews and demographic questionnaires. We included participants that actively practiced as licensed RDs and were ≥ 22 years of age (n=20). Interviews were conducted over Zoom. RD perspectives were recorded through audio recordings that were professionally transcribed. Transcriptions were inductively analyzed and codified through Dedoose to generate common themes. Qualitative themes included significance of current methods of dietary recall, barriers and facilitators of current methods of dietary recall, and barriers and facilitators of VAS dietary recall.

Results
Of the 20 interviewees, ages ranged from 22-65+ years, of which 9 (45%) were 22-34, 8 (40%) were 35-64, and 3 (15%) were ≥ 65 years. Most of the RDs identified as being female (95%), white (85%), and not of Hispanic, Latino/a/x, or Spanish origin (90%). RDs believed the implementation of VAS had the potential to improve dietary recall assessments of older adults. RDs reported challenges of current dietary recall methods among older adults, such as memory and hearing (65% of participants) and properly reporting serving sizes (45% of participants). RDs reported that VAS dietary recall may be beneficial due to convenience of real-time data collection (80% of participants), more efficient use of time during clinical visits (65% of participants), and ease of accurately reporting meals (85% of participants). However, RDs reported low technology literacy in older adults as a barrier to VAS implementation (70% of participants).

Conclusion
Our qualitative study provides a framework on how VAS can better meet the needs of both older adults and clinical RDs in managing and assessing dietary intake. Future VAS technologies can improve dietary recall methods by incorporating probing cues to enhance real- time data collection, providing clinicians accessibility to patient responses, and including questions that enable collection of data points related to food intake, such as accurate intake times and food-related symptoms.

Understanding the Feasibility of Voice-assistant Systems to Advance Dietary Recall Methods

Association of Strict Versus Lenient Cholesterol Lowering with Cardiac Outcomes, Diabetes Progression and Complications, and Mortality in Patients with Diabetes Treated with Statins: Is Less More?

Background: Accurately predicting the healing time of wounds is a critical aspect of wound care. Developing models that can predict when a wound will heal based on different parameters can provide the healthcare provider and the patient with valuable information. This knowledge can help tailor personalized treatment plans, and optimize resource allocation, thus enhancing patient outcomes. The current project aims to create a predictive model capable of providing precise estimates of wound healing time. 

Methods: A retrospective analysis of 1061 healed wounds was performed using Tissue Analytics (TA). The wounds were stratified into one of six wound etiologies: diabetic, arterial, venous, pressure, traumatic and idiopathic. Different metrics such as total area, length, width and perimeter of the wound, red tissue, black tissue, yellow tissue, other tissue and pink tissue percentages were extrapolated. The number of days in treatment was calculated by subtracting the first presenting day by the day the wound healed. The above information was compiled into a CSV file and Google Colaboratory was employed to generate different regression models using Pycaret. Pycaret automatically tests different machine learning regression models and their efficacy in predicting wound heal time in days.

Results: 20 different machine learning regression models were used and compared to analyze the data and make predictions. Of the 20 different machine learning models, the Random Forest Regressor (RF) had the lowest mean squared error of 1109.46 and root mean squared error of 33.25 days. Extra Trees Regressor (ET) model had the lowest mean absolute error of 26.81 days and mean absolute percentage of 35.51%. The RF model also had the best R-squared of 0.12. A feature importance plot was also generated, which showed that red tissue percentage and etiology were the greatest predictors of wound heal time.

Conclusion: The results showed a high rate of error and low accuracy as indicated by the relatively high MSE and RMSE as well as a low R-sqared for almost all models. Extra metrics generated by Pycaret showed the presence of overfitting, meaning that the models memorized the different numbers to make predictions instead of evaluating the different features. Prior research has shown that wound perimeter and wound area are more likely to predict heal time; however, the data structure could force the the model to determine that etiology and red tissue is best predictor of wound heal time. The current study is limited by the number of wounds as well as other variables that can affect wound heal time such as compliance; however, it can be improved by increasing the number of wounds to train the model.

Premium content

Evaluating the diagnostic utility of ChatGPT for providing dermatopathology immunohistochemistry reference recommendations

Next from 2023 AMA Research Challenge

Cutaneous Neoplasms as Donors and Recipients in Tumor-to-Tumor Metastasis: A Systematic Review of the Literature

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES