Precision And Recall Research Articles

Clinical large language model to predict loss to follow up for oncology patients discharged to skilled nursing facilities.

e13597 Background: From SEER data, there is evidence that oncology patients who are discharged to a skilled nursing facility are less likely to receive chemotherapy and radiation. Those that are unlikely to follow up are a vulnerable population that should be identified to discuss how to best optimize post hospitalization care. Methods: Through University of Florida (UF)’s Integrated Data Repository of EHR data, we identified 591 hospitalization events for UF established oncology patients that lead to a discharge to a skilled nursing facility from UF/Shands Hospital between 2012 – 2023. The outcome of interest was loss to follow up, defined as no clinic follow up within 6 months after discharge. 222 hospitalization events lead to loss to follow up and 369 did not. We trained XGBoost, random forest and logistic regression models on structured EHR data such as ICD codes, patient demographics including insurance payer, Charlson Comorbidity Index, clinical features such as discharge vitals with a data split of 6:4 for training and testing to predict loss to follow up. We compared this against GatorTron, a clinical large language model, which was trained separately on hospital discharge summaries and concatenated notes from physical therapists, occupational therapists, case managers and social workers with a data split of 6:2:2 for training, validation and testing. Results: As noted in the table, XGBoost performed better than logistic regression or random forest with a F1 score of 0.298. GatorTron_345m using physician notes from hospital discharge summaries, however yielded a poorer performance with a F1 score of 0.229. With additional notes from physical therapists and occupational therapists, case managers and social workers, GatorTron_345m achieved the best performance amongst all the models with a F1 score of 0.377 using a learning rate of 1e-5 for 30 epochs. Conclusions: To the best of our knowledge, there is no study leveraging machine learning, especially large language models to predict loss to follow up in oncology clinics. This is the first study to initiate predictive modelling with respect to this important clinical question. This study also highlights the value of multidisciplinary notes in predicting post-hospitalization outcomes rather than physician notes only. Loss to follow up for oncology patients is a multifactorial issue and despite the collection of a broad selection of patient and encounter features, as well as the employment of advanced machine learning tools in this study, elucidation of better determinants is needed. Model performance results. Input Model F1 Precision (PPV) Recall (sensit) Specificity Accuracy ICD + demographics XGBoost 0.298 0.362 0.254 0.729 0.550 Random Forest 0.137 0.3 0.089 0.873 0.578 LogisticRegression 0.264 0.3 0.236 0.669 0.506 Discharge Summaries GatorTron_345m 0.229 0.444 0.154 0.882 0.606 Multidisciplinary Notes 0.377 0.482 0.310 0.825 0.648

Enhancing personalized prognosis in appendiceal (AC) and cecal adenocarcinoma (CA) with machine learning.

e16484 Background: AA and CA are rare gastrointestinal neoplasms (2% and 1.2%, respectively) that are understudied and difficult to diagnose. We aimed to apply machine learning (ML) to identify prognostic factors and to enhance survival outcomes. Methods: Data were obtained from the SEER database (2004-2021). Patients who met any of the following criteria were excluded: diagnosis not confirmed by histology, previous history of cancer or other concurrent malignancies, or unknown data. To identify prognostic variables, we conducted Cox regression analysis and constructed prognostic models using ML algorithms to predict the 5-year survival. Patient records were randomly divided into training (70%) and validation (30%) sets. A validation method incorporating the area under the curve (AUC) of the receiver operating characteristic curve was used to validate the accuracy and reliability of the ML models. We also investigated the role of multiple therapeutic options using Kaplan-Meier survival analysis. Results: In total, 136,010 patients were included. Of these, 120,000 had CA and 16,010 had AC. Most patients (56%) were female and 80.3% were White. The median patient age was 69 years, and the median tumor size was 4.4 cm. Most patients were T1 (54.2%) and T4 was rare (16.5%). Nodal involvement was observed in 14.7% of cases, while M0 accounted for 91.7%. Of the metastatic cases, 1.2% were in the liver and 0.2% in the liver and lungs. Among the patients, 99.4% underwent surgery, 0.5% received radiation, and 15.7% received chemotherapy. The 5-year overall survival (OS) was higher in patients aged < 69 years (68.8%) than in those aged ≥ 69 years (48.4%). Survival also varied by primary site, with appendix tumors showing a significantly higher OS (82.1%) than cecal tumors (55.7%). Metastatic status strongly influenced OS, with M0 having a much higher rate (67.5%) than M1 (10.9%). Similarly, nodal involvement impacted outcomes, as N0 patients had the highest OS (72.7%), followed by N1 (49.6%), and N2 (25.5%). Multivariate Cox regression analysis identified older age, male sex, radiation, metastasis, and White and Black race as poor prognostic factors, whereas surgery and chemotherapy were good prognostic factors. Gradient boosting and Random Forest were the most accurate models. The ML models identified TNM stage as the most significant prognostic factor, followed by age and primary site. The performance metrics for all ML algorithms are summarized in Table. Conclusions: ML plays a crucial role in personalized medicine by identifying the key prognostic factors in AC and CA. Further research is needed to assess the benefits of radiation therapy for managing these rare malignancies. ML Model Accuracy Precision Recall F1 score AUC LR 64.9% 62.2% 38% 47.1% 0.716 KNN 61.9% 52.7% 74.2% 61.6% 0.664 RFC 65.1% 57.6% 58.1% 57.8% 0.722 GBC 65% 57.9% 55.% 56.4% 0.719 MLP 64% 57.5% 57.3% 57.4% 0.719

Predicting survivorship and caregiver burden in multiple myeloma (MM) using machine learning (ML): Insights from a large, national, prospective study.

e13661 Background: MM survival has improved significantly, resulting in increased survivorship and caregiver burden and is crucial to understand the real-world impact of MM. We analyzed results from a large, prospective MM patient survey using ML models to identify key predictive factors for survivorship and caregiver burden. Methods: A 45-item questionnaire spanning sociodemographics, treatment preferences, symptom burden and caregiver impact was distributed to MM patients at Mayo Clinic and International Myeloma Foundation support groups. Questions about quality of life (QOL) and treatment satisfaction (TS) represented survivorship outcomes while caregiver workday loss (WL) and job loss (JL) were surrogates for caregiver burden. 6 ML models (multinomial, random forest (RF), XGB (XGBoost), k-NN, GBM, Naïve Bayes) with a training: testing cohort of 80:20 were used to predict each of the 4 outcomes independently, considering model accuracy, C-index, precision, recall, and F1 score (Table), and using R studio (V4.4.1) and Bluesky Statistics (V10.3.4). Results: 2,239 participants (54% male, 89% Whites) with median age 68 yrs completed the survey. 71% respondents had a college degree, 44% resided in Midwest states, 62% reported household income >$75,000, and 67% had Medicare/Medicaid. Median age at MM diagnosis was 67 yrs with 58% diagnosed within the past 5 yrs and 73% currently receiving MM treatment. 72% reported their spouse as the primary caregiver. XGB was the best predictive model for QOL with gender, household income, stem cell transplant, CAR-T therapy, support from medical staff, and symptom burden (pain, neuropathy, mental well-being) as significant predictors. The RF model was the best for TS with pain and neuropathy levels as the most impactful predictors. For caregiver burden, XGB was the best model for WL with gender, household income, CAR-T therapy, financial strain, lack of support from medical staff, self-workdays missed, self-job loss, and pain and neuropathy levels as significant predictors. For JL, RF was the best model with significant variables being workdays missed by self, increased healthcare needs, symptom burden (pain, neuropathy, and mental well-being), and restricted physical activity. Conclusions: Our study shows the significant impact of MM survivorship and efficiently identifies predictive factors associated with survivorship and caregiver burden from a large prospective dataset using validated ML models. These findings highlight the complex challenges from MM and provide contemporary ML methods to better understand their interplay such that a framework of real-world interventions can be developed. Outcome Best ML Model Accuracy C-Index Precision Recall F1 score QOL XG 0.996 0.996 0.996 0.997 0.996 TS RF 0.984 0.900 0.937 0.948 0.937 WL XGB 0.984 1.000 0.944 0.947 0.945 JL RF 0.989 0.800 0.829 0.848 0.835

Prediction of sotorasib response duration and identification of clinically actionable resistance mechanisms using an RNA- and machine learning–based patient classifier.

e13614 Background: KRAS mutations are the most common oncogenic drivers in cancer. While first-generation KRAS inhibitors (KRASi) represent a therapeutic advancement, objective response rates and therapeutic durability remain suboptimal, with the majority of patients progressing on treatment. The mechanisms of resistance to KRASi therapy and clinical strategies for follow-on treatments are not well-defined. Methods: 72 biopsies from non-small cell lung cancer (NSCLC) patients with known time on treatment with sotorasib were analyzed by RNA-sequencing and the krasID biomarker model. Individual biological module scores served as input features for a classification model. Tuning and model refinement was performed with nested cross-validation. 5 NSCLC patients with pre- and post-sotorasib biopsies were evaluated for longitudinal changes in tumor biology as assessed by krasID modules. Results: This study comprised longitudinal evaluation of real-world NSCLC patients treated with sotorasib, a first-generation KRASi, analyzed with a RNA- and machine learning-based patient classifier, Genialis™ krasID. The krasID classifier stratified patients by scoring a number of biological signatures (“modules”) relevant to KRAS-mediated oncogenesis, and predicted therapy responses by algorithmic integration of these signatures. For KRASi-specific stratification, the classifier achieved an AUROC of 0.81 and specificity of 0.86 in predicting KRASi response. Beyond predicting initial drug response, we investigated the functionality of the biomarker to inform whether patients remained on sotorasib treatment at 6 months (AUROC > 0.8, see Table), a time point aligning with the median progression-free survival of 5.6 months from CodeBreaK 200 trial. The response-duration classifier proved capable of predicting the time to disease progression, an important clinical parameter considering that approximately 2/3 of patients progress on sotorasib by 6 months. Analysis of pre- and post-treatment biopsies in a subset of these patients implicated shared and unique resistance mechanisms. krasID identified certain patients with compensatory resistance mechanisms outside the MAPK/KRAS signaling cascade who nevertheless continued to show molecular benefit to sotorasib. Patient tolerable combination therapies targeting secondary resistance mechanisms, alongside continued sotorasib treatment, may improve outcomes compared to discontinuation of KRASi. Conclusions: These findings position krasID as a next-generation biomarker for identifying KRASi responders, predicting treatment duration, and guiding personalized follow-on therapies for KRAS-mutated cancers. Performance of time on sotorasib classifier. Accuracy Precision Recall Specificity AUROC Classifier 0.78 0.78 0.69 0.85 0.80 Dummy model 0.51 0.44 0.42 0.58 0.48

Development and validation of machine learning risk prediction models for detection of early-onset colorectal cancer: Data from 30 health systems in the United States.

3613 Background: Incidence of early-onset colorectal cancer (EoCRC) in patients without any family history has been increasing in recent years. Our study leveraged advanced Large Language Models (LLM), like GPT-4, to predict EoCRC in a population comprised of multiple health systems across the United States. There is potential to improve patient care by suggesting early screening for patients who are predicted to be at risk by the model. Methods: We identified a population of 5532 patients aged between 18 and 44 in the Truveta data, which is a collection of 120+ million patient journeys across 30 U.S. health systems. 1376 (24.87%) were diagnosed of CRC based on their ICD and SNOMED-CT codes. Data was split into training (80%) and testing (20%) sets. For the prediction task, we applied GPT-4o and compared with XGBoost, one of the strongest non-generative machine learning models. We used patient demographics (age, gender, race, ethnicity), conditions, and lab results within the last 2 to 7 months prior to CRC diagnosis for model training. The last month before CRC diagnosis was excluded to avoid highly predictive signals. For XGBoost, the demographics were represented as one-hot feature vectors, indicating their presence or absence. Conditions were encoded by their diagnosis frequency within the time frame, while the actual values of the lab results were used as model features. For the LLM model, all patient information was input as plain text. Both conditions and lab results were represented by the names of ICD and LOINC codes in order to capture the clinical context. A Chain-of-Thought prompting strategy, incorporating detailed instructions and CRC-specific knowledge, was employed to guide the LLM. Results: Our test set consisted of 1105 patients in which 279 (25.25%) were diagnosed with CRC. Both XGBoost and GPT-4o achieved comparable results. Despite the uneven distribution of CRC diagnoses in the test samples, the fine-tuned GPT-4o achieved the highest precision (87.43%) and recall (57.35%). This indicates that the model can accurately predict EoCRC in the near future for over half of the patients. Conclusions: This study highlights the potential of using LLM to predict EoCRC in younger population. While the GPT-4o base model contains general medical knowledge, supervised fine-tuning with explicit guideline enhances its predictive capabilities. The high precision of the model performance minimizes the burden of unnecessary screening by identifying patients with relatively high risk. As AI technologies continues to advance, with sufficient governance policy in place, predictive models can be valuable tools for clinicians to suggest early screening and mitigate EoCRC for patients. Model Precision Recall F1 score Accuracy XGBoost 83.77% 57.35% 68.09% 86.43% GPT-4o (base) 68.61% 54.84% 60.69% 82.26% GPT-4o (fine-tuned) 87.43% 57.35% 69.26% 87.15%

Identifying key prognostic indicators in Wilms tumor using machine learning techniques.

4557 Background: Wilms tumor is a rare pediatric malignancy, accounting for 6% of pediatric tumors and primarily affecting the kidneys. Its impact on quality of life and long-term outcomes complicates management. This study leveraged machine learning (ML) to identify prognostic factors with the aim of enhancing prognosis and survival rates. Methods: Data were obtained from the SEER database (2004-2021). Patients who met any of the following criteria were excluded: diagnosis not confirmed by histology, previous history of cancer or other concurrent malignancies, or unknown data. To identify prognostic variables, we conducted Cox regression analysis and constructed prognostic models using ML algorithms to predict the 5-year survival. Patient records were randomly divided into training (70%) and validation (30%) sets. A validation method incorporating the area under the curve (AUC) of the receiver operating characteristic curve was used to validate the accuracy and reliability of the ML models. We also investigated the role of multiple therapeutic options using Kaplan-Meier survival analysis. Results: A total of 4,935 children were included. Among them, 47.72% underwent surgery, radiation, and chemotherapy; 45.07% underwent surgery and chemotherapy; and 7.21% underwent surgery alone. Most patients (53.3%) were females, 75% were white, followed by black (16.3%). The mean patient age was 3 years, and the mean tumor size was 10.6 cm. Most tumors were left-sided (51.1%) and 79.8% had no metastasis. The lungs were the most frequent site of metastasis (11%), followed by the liver and lungs at the same time (1.2%), and bone involvement was rare (0.6%). Radical surgery was the most common surgical approach (76.1%), followed by nephrectomy (4.9%). Patients who underwent surgery and chemotherapy had the highest 5-year OS (96.9%) and CSS (96.9%) compared to those who underwent surgery alone (OS: 95.5%, CSS: 95.5%) or surgery with chemotherapy and radiation (OS: 92.9%, CSS: 93.2%). Asian/Pacific Islander and white patients exhibited better OS (94.3% and 95.1%, respectively) than black patients (91.9%). Multivariate Cox regression analysis identified a large tumor size and older age as poor prognostic factors. Gradient boosting and MLP classifiers were the most accurate models. The ML models identified race as the most significant prognostic factor, followed by the TNM stage and age. The performance metrics for all ML algorithms are summarized in Table. Conclusions: This is the first study to apply ML to Wilms tumor, effectively identifying key prognostic factors. ML models show promise in enhancing survival predictions, potentially informing personalized treatment strategies, and improving patient outcomes. ML Model Accuracy Precision Recall F1 score AUC LR 59.5% 59.5% 99.6% 74.5% 0.573 KNN 56.1% 61.7% 68.7% 65.05% 0.572 RFC 59.9% 61.9% 84.4% 71.4% 0.596 GBC 60.7% 60.5% 97.3% 74.6% 0.582 MLP 60.9% 61.6% 90.5% 73.3% 0.589

Use of large language models to extract cancer diagnosis, histology, grade, and staging from unstructured electronic health records.

11176 Background: The oncology ecosystem contains millions of unstructured documents within Electronic Health Record (EHR) systems, including clinical notes and pathology reports with vital patient information like cancer diagnoses, histology, grade, and staging. These documents are often text or scanned PDFs. Extracting clinical details from these sources can improve EHR completeness and accuracy. To address this challenge, we developed an information extraction pipeline that leverages the recent advances in artificial intelligence (AI). Methods: Large Language Models (LLMs) and prompt engineering was used to extract information from both clinical notes and pathology reports. For pathology reports, Optical Character Recognition (OCR) was applied to convert scanned images into text, which was then processed by the LLM using a tailored prompt designed to extract the relevant cancer diagnosis and staging details. For clinical notes, the text was directly passed into the LLM along with an optimized prompt to extract the same clinical information. The information extraction pipeline was validated using a dataset of 829 pathology reports and 569 progress notes from the EHR system across 40 cancer types, including 26 solid tumors and 14 hematologic malignancies. Clinical specialists manually extracted cancer diagnoses, histology, grade, and staging, which were compared to the automated output. An F1 score using a combined measure of precision and recall was calculated to assess the model’s accuracy. Results: The pipeline retrieved cancer type, histology, and grade from pathology reports with an F1 score over 0.85. Similarly, it extracted cancer type and staging information from progress notes with an F1 score over 0.85, demonstrating high accuracy and reliability. Additional performance metrics are shown in the table. Conclusions: This LLM-based extraction pipeline accurately identified cancer diagnoses, histology, grade, and staging information from unstructured text and scanned documents within the EHR achieving an F1 score ≥0.85. The validation results suggest that this approach could be scaled to improve the completeness and utility of EHR data, supporting the availability of more robust information for scientific research, clinical care, and other uses of health data. Performance metrics of LLM-based extraction pipeline. Information Source Precision Recall Specificity NPV* Accuracy F1 > Cancer Type Pathology (PDF) 0.86 0.89 0.99 0.98 0.89 0.87 Histology Pathology (PDF) 0.83 0.88 0.90 0.88 0.90 0.85 Grade Pathology (PDF) 0.8 0.95 0.96 0.96 0.88 0.87 Cancer Type Progress Notes 0.8 0.95 0.96 0.96 0.88 0.87 Staging TNM (T) Progress Notes 0.96 0.87 0.99 0.97 0.97 0.91 Staging TNM (N) Progress Notes 0.92 0.85 0.99 0.97 0.96 0.88 Staging TNM (M) Progress Notes 0.87 0.83 0.98 0.97 0.96 0.85 *NPV= negative predictive value.

Fine-grained entity disambiguation through numeric pattern awareness in transformer models

Knowledge base question answering systems rely on entity linking to connect textual mentions in natural language with corresponding entities in a structured knowledge base. While conventional methods perform well in static or generic scenarios, they often struggle in dynamic contexts that require fine-grained disambiguation between specific instances. This limitation is particularly evident in domains such as military situation reports, where accurate identification of individual instances is essential for timely and precise information retrieval. This work introduces a novel framework for instance-level entity linking that enables transformer-based language models to distinguish between similarly described but distinct instances of the same entity. The proposed framework integrates three core innovations: (1) a context-aware masking strategy that reduces over-reliance on surface-level entity mentions by encouraging the model to leverage surrounding contextual cues, (2) a multi-task learning component that improves the model’s understanding of relative spatial and temporal information, and (3) a position-aware representation technique that restructures numerical data into token sequences that preserve structural integrity, addressing inherent limitations of subword tokenization in current language models. Experimental evaluations on simulated Korean military reports demonstrate substantial performance gains, particularly in challenging cases involving ambiguous or overlapping entity descriptions. The enhanced model shows a markedly improved ability to differentiate between instances based on subtle temporal and geospatial cues. These advancements extend beyond entity linking, offering broader insights into how transformer-based models can more effectively interpret numerical information. This research establishes a new direction for precise entity disambiguation in dynamic environments and contributes practical methods for enhancing language model performance across critical domains, including medical diagnostics, financial analysis, scientific research, autonomous systems, and numerical reasoning in smaller Large Language Models, where our techniques can potentially bridge the performance gap with larger models without requiring extensive computational resources.

An integrative and comprehensive analysis of blood transcriptomes combined with machine learning models reveals key signatures for tuberculosis diagnosis and risk stratification

Tuberculosis (TB) remains a major global health challenge, contributing substantially to morbidity and mortality worldwide. The progression from Mycobacterium tuberculosis (Mtb) infection to active disease involves a complex interplay between host immune responses and Mtb’s ability to evade them. However, current diagnostic tools, such as interferon-gamma release assays (IGRAs) and tuberculin skin tests (TSTs), have limited ability to distinguish between different stages of TB or to predict the progression from infection to active disease. In this study, we performed an integrative analysis of 324 previously acquired blood transcriptome samples from TB patients, TB contacts, and controls across diverse geographical regions. Differential gene expression analysis revealed distinct transcriptomic signatures in TB patients, highlighting dysregulated pathways related to immune responses, antimicrobial peptides, and extracellular matrix organization. Using machine learning, we identified a 99-transcript signature that accurately distinguished TB patients from controls, demonstrated strong predictive performance across different cohorts, and identified potential progressors or subclinical cases. Validation in an independent dataset comprising 90 TB patients and 20 healthy controls confirmed the robustness of the 10-gene signature (BATF2, FAM20A, FBLN2, AK5, VAMP5, MMP8, KLHDC8B, LINC00402, DEFA3, and GBP6), achieving high area under the curve (AUC) values in both receiver operating characteristic (ROC) and precision–recall analyses. This 10-gene signature offers promising candidates for further validation and the development of diagnostic and prognostic tools, supporting global efforts to improve TB detection and risk stratification.

Differential Profiles of Gut Microbiota-Derived Metabolites of Bile Acids and Propionate as Potential Predictors of Depressive Disorder in Women with Morbid Obesity at High Risk of Metabolic Dysfunction-Associated Steatotic Liver Disease—A Pilot Study

Metabolic dysfunction-associated steatotic liver disease (MASLD) is a liver condition linked to cardiometabolic diseases and mental health issues, with studies highlighting disruptions in gut microbiota activity, including bile acid (BA) metabolism. Therefore, the main aim of this exploratory analysis was to assess microbiota-derived metabolites, specifically BAs and short-chain fatty acids (SCFAs), as potential biomarkers of depressive disorder (DD) in women with morbid obesity at MASLD risk. In this pilot study, 33 females with morbid obesity who were scheduled for bariatric surgery were evaluated. Medical and clinical data were collected, and microbial metabolites from pre-surgery blood samples were analyzed. Patients were stratified according to the presence of DD. Analysis with Spearman’s rank test was used to assess correlations and logistic regression models were built to evaluate biomarkers as predictors of DD risk using both receiver operating characteristic (ROC) and precision–recall curves. In this cohort, 30.3% of females were reported to have DD, in addition to significantly elevated levels of certain BAs and SCFAs, including glycodeoxycholic acid (GDCA) and propionate, which were also correlated with some metabolic biomarkers. However, there were no differences in the incidence of MASLD or metabolic syndrome between patients with DD or without. In conclusion, microbiota-derived metabolites such as GDCA and propionate may influence DD risk in females with morbid obesity; however, their potential use as predictive biomarkers should be further investigated to confirm their role in psycho-metabolic conditions.

Research and application of deep learning object detection methods for forest fire smoke recognition

Forest fires are severe ecological disasters worldwide that cause extensive ecological destruction and economic losses while threatening biodiversity and human safety. With the escalation of climate change, the frequency and intensity of forest fires are increasing annually, underscoring the urgent need for effective monitoring and early warning systems. This study investigates the application effectiveness of deep learning-based object detection technology in forest fire smoke recognition by using the YOLOv11x algorithm to develop an efficient fire detection model. The objective is to enhance early fire detection capabilities and mitigate potential damage. To improve the model’s applicability and generalizability, two publicly available fire image datasets, WD (Wildfire Dataset) and FFS (Forest Fire Smoke), encompassing various complex scenarios and external conditions, were employed. After 501 training epochs, the model’s detection performance was comprehensively evaluated via multiple metrics, including precision, recall, and mean average precision (mAP50 and mAP50-95). The results demonstrate that YOLOv11x excels in bounding box loss (box loss), classification loss (cls loss), and distribution focal loss (dfl loss), indicating effective optimization of object detection performance across multiple dimensions. Specifically, the model achieved a precision of 0.949, a recall of 0.850, an mAP50 of 0.901, and an mAP50-95 of 0.786, highlighting its high detection accuracy and stability. Analysis of the precision‒recall (PR) curve revealed an average mAP@0.5 of 0.901, further confirming the effectiveness of YOLOv11x in fire smoke detection. Notably, the mAP@0.5 for the smoke category reached 0.962, whereas for the flame category, it was 0.841, indicating superior performance in smoke detection compared with flame detection. This disparity primarily arises from the distinct visual characteristics of flames and smoke; flames possess more vivid colors and defined shapes, facilitating easier recognition by the model, whereas smoke exhibits more ambiguous and variable textures and shapes, increasing detection difficulty. In the test set, 86.89% of the samples had confidence scores exceeding 0.85, further validating the model’s reliability. In summary, the YOLOv11x algorithm demonstrates excellent performance and broad application potential in forest fire smoke recognition, providing robust technical support for early fire warning systems and offering valuable insights for the design of intelligent monitoring systems in related fields.

Predicting Agitation Events in the Emergency Department Through Artificial Intelligence

Agitation events are increasing in emergency departments (EDs), exacerbating safety risks for patients and clinicians. A wide range of clinical etiologies and behavioral patterns in the emergency setting make agitation prediction difficult in this setting. To develop, train, and validate an agitation-specific prediction model based on a large, diverse set of past ED visit data. This cohort study included electronic health record data collected from 9 ED sites within a large, urban health system in the Northeast US. All ED visits featuring patients aged 18 years or older from January 1, 2015, to December 31, 2022, were included in the analysis and modeling. Data analysis occurred between May 2023 and September 2024. Variables that served as potential exposures of interest, encompassing demographic information, patient history, initial vital signs, visit information, mode of arrival, and health services utilization. The primary outcome of agitation was defined as the presence of an intramuscular chemical sedation and/or violent physical restraint order during an ED visit. A clinical model was developed to identify risk factors that predict agitation development during an ED visit prior to symptom onset. Model performance was measured using area under the receiver operating characteristic curve (AUROC) and area under the precision recall curve (PR-AUC). The final cohort comprised 3 048 780 visits. The cohort had a mean (SD) age of 50.2 (20.4) years, with 54.7% visits among female patients. The final artificial intelligence model used 50 predictors for the primary outcome of predicting agitation events. The model achieved an AUROC of 0.94 (95% CI, 0.93-0.94) and a PR-AUC of 0.41 (95% CI, 0.40-0.42) in cross-validation, indicating good discriminative ability. Calibration of the model was evaluated and demonstrated robustness across the range of predicted probabilities. The top predictors in the final model included factors such as number of past ED visits, initial vital signs, medical history, chief concern, and number of previous sedation and restraint events. Using a cross-sectional cohort of ED visits across 9 hospitals, the prediction model included factors for detecting risk of agitation that demonstrated high accuracy and applicability across diverse patient populations. These results suggest that clinical application of the model may enhance patient-centered care through preemptive deescalation and prevention of agitation.

Diagnosis and risk stratification of PI-RADS v2.1 category 3–5 lesions using amide proton transfer imaging

ObjectivesTo explore the value of amide proton transfer-weighted imaging (APTWI) combined with apparent diffusion coefficient (ADC) and prostate-specific antigen density (PSAD) in evaluating the risk of aggressive PI-RADS v2.1 category 3–5 lesions.Materials and methodsWe retrospectively analyzed clinical and MRI features of 69 patients with prostate cancer (PCa) and 32 patients with benign prostatic lesion (BL). The PCa group was classified into Gleason Grade Groups (GG) 1 to 4 based on Gleason Score (GS). APTWI parameters, ADC, and PSAD were compared among the different groups. The receiver operating characteristic (ROC) curve and precision recall (PR) curve were used to assess the diagnostic accuracy of each parameter, and Spearman rank correlation was used to analyze the relationship between APTWI parameters and GS.ResultsAPTWI parameters and PSAD were significantly higher in the PCa group than that in the BL group, whereas ADC was significantly lower in the PCa group than in the BL group. ADC showed the highest AUC in the diagnosis of PCa, followed by PSAD and APTmin. Combined analysis showed that APTmin+ADC + PSAD exhibited the highest AUC (0.997). In the PCa group, significant differences in APTWI parameters were found among GG1 to GG4 (P < 0.001), with intra-group comparisons showing significant differences between GG1 and GG3, GG1 and GG4, GG2 and GG3, and GG2 and GG4 subgroups. The AUC of APTmean was greatest in evaluating the risk of aggressive PCa (0.948), which further increased when APTmean was combined with ADC and PSAD (0.956).ConclusionIn PI-RADS v2.1 category 3–5 lesions, APT can serve as an important biomarker for the risk stratification of PCa, and combining APT with PSAD and ADC achieves the highest diagnostic efficacy.

Noise-Robust image edge detection based on multi-scale automatic anisotropic morphological Gaussian Kernels.

This paper presents a novel multi-scale, noise-robust edge detection method that employs multi-scale automatic anisotropic morphological Gaussian kernels to extract edge maps from input images. It addresses the issue of cross-edge detection failure in the Canny edge detector. Compared to other edge detection methods, the proposed approach offers significant advantages in maintaining noise robustness while achieving high edge resolution and accuracy. The paper is structured into five key sections. First, we propose a multi-scale automatic anisotropic morphological directional derivative (AMDD) to capture local gray-level variations around each pixel at multiple scales. Second, a new fused edge strength map (ESM) is introduced based on the multi-scale AMDD. Third, we analyze why the Canny isotropic Gaussian kernel detector fails to detect cross edges. Additionally, the edge contour is extracted by incorporating the fused ESMs and the edge direction map (EDM), which are processed through spatial and directional matching filters, into the standard Canny detection framework. Finally, we evaluate the proposed method using precision-recall (PR) curves and Pratt's Figure of Merit (FOM). We compare its performance with existing state-of-the-art detectors on a standard dataset. Experimental results demonstrate that the proposed method effectively reduces noise, mitigates irrelevant signal interference, and smooths the image, showing competitive performance in edge detection tasks.

Explainable machine learning model for prediction of 28-day all-cause mortality in immunocompromised patients in the intensive care unit: a retrospective cohort study based on MIMIC-IV database

ObjectivesThis study aimed to develop and validate an explainable machine learning (ML) model to predict 28-day all-cause mortality in immunocompromised patients admitted to the intensive care unit (ICU). Accurate and interpretable mortality prediction is crucial for clinical decision-making and optimal allocation of critical care resources for this vulnerable patient population.MethodsWe utilized retrospective clinical data from the MIMIC-IV (version 2.2) database, encompassing ICU admissions at Beth Israel Deaconess Medical Center from 2008 to 2019. Eligible immunocompromised patients, including those with primary immunodeficiencies and chronic acquired conditions, such as hematological malignancies, solid tumors, and organ transplantation, were selected. Data were randomly split into training (80%) and testing (20%) cohorts. Ten ML models (logistic regression, XGBoost, LightGBM, AdaBoost, Random Forest, Gradient Boosting, Gaussian Naive Bayes, Complement Naive Bayes, Multilayer Perceptron, and Support Vector Machine) were developed and evaluated using area under the receiver operating characteristic curve (AUROC), area under the precision–recall curve (AUPRC), sensitivity, specificity, accuracy, and F1 score. Model explainability was achieved through SHapley Additive exPlanations (SHAP), and decision curve analysis (DCA) assessed clinical utility. In addition, Cox proportional hazards regression was conducted to evaluate the impact of predictive factors on time-to-event outcomes.ResultsAmong the evaluated models, the Support Vector Machine (SVM) demonstrated the highest AUROC of 0.863 (95% CI 0.834–0.890) and a notable AUPRC of 0.678 (95% CI 0.624–0.736). Key predictive factors consistently identified across multiple ML models included 24-h urine output, blood urea nitrogen (BUN) levels, presence of metastatic solid tumors, Charlson Comorbidity Index (CCI), and international normalized ratio (INR). SHAP analyses provided detailed insights into how these features influenced model predictions.ConclusionsThe explainable ML models based on various artificial intelligence methods demonstrated promising clinical applicability in predicting 28-day mortality risk among immunocompromised ICU patients. Factors such as urine output, BUN, metastatic solid tumors, CCI, and INR significantly contributed to prediction outcomes and may serve as important predictors in clinical practice.

High-Granularity Machine Learning Prediction of Acute Brain Injury in Patients Receiving Venoarterial Extracorporeal Membrane Oxygenation.

Acute brain injury (ABI) is prevalent among patients undergoing venoarterial extracorporeal membrane oxygenation (VA-ECMO) and significantly impact recovery. Early prediction of ABI could enable timely interventions to prevent adverse outcomes, but existing predictive methods remain suboptimal. This study aimed to enhance ABI prediction using machine learning (ML) models and high-temporal-resolution granular data. We retrospectively analyzed 355 VA-ECMO patients treated at Johns Hopkins Hospital (JHH) from 2016 to 2024, collecting over 3 million data points from the JHH Research Electronic Data Capture (REDCap) database, with an average of 80,000 data points per patient. Acute brain injury was defined as ischemic stroke, intracranial hemorrhage, hypoxic-ischemic brain injury, or seizure. Four ML models were used: Random Forest, Categorical Boosting, Adaptive Boosting, and Extreme Gradient Boosting. Among 355 patients (median age 59 years, 56.9% male), 13.5% developed ABI. The models achieved an optimal area under the receiver operating characteristic curve (AUROC) of 0.79, accuracy of 87%, sensitivity of 53%, specificity of 99%, and precision-recall (PR)-AUC of 0.47. Key predictors included high minimum values of systolic blood pressure and variability in on-ECMO pulse pressure. High-resolution granular data enhanced ML performance for ABI prediction. Future efforts should focus on integrating continuous data platforms to enable real-time monitoring and personalized care, optimizing patient outcomes.

Intersectional and Marginal Debiasing in Prediction Models for Emergency Admissions

Fair clinical prediction models are crucial for achieving equitable health outcomes. Intersectionality has been applied to develop algorithms that address discrimination among intersections of protected attributes (eg, Black women rather than Black persons or women separately), yet most fair algorithms default to marginal debiasing, optimizing performance across simplified patient subgroups. To assess the extent to which simplifying patient subgroups during training is associated with intersectional subgroup performance in emergency department (ED) admission models. This prognostic study of admission prediction models used retrospective data from ED visits to Beth Israel Deaconess Medical Center Medical Information Mart for Intensive Care IV (MIMIC-IV; n = 160 016) from January 1, 2011, to December 31, 2019, and Boston Children's Hospital (BCH; n = 22 222) from June 1 through August 13, 2019. Statistical analysis was conducted from January 2022 to August 2024. The primary outcome was admission to an in-patient service. The accuracy of admission predictions among intersectional subgroups was measured under variations on model training with respect to optimizing for group level performance. Under different fairness definitions (calibration, error rate balance) and modeling methods (linear, nonlinear), overall performance and subgroup performance of marginal debiasing approaches were compared with intersectional debiasing approaches. Subgroups were defined by self-reported race and ethnicity and gender. Measures include area under the receiver operator characteristic curve (AUROC), area under the precision recall curve, subgroup calibration error, and false-negative rates. The MIMIC-IV cohort included 160 016 visits (mean [SD] age, 53.0 [19.3] years; 57.4% female patients; 0.3% American Indian or Alaska Native patients, 3.7% Asian patients, 26.2% Black patients, 10.0% Hispanic or Latino patients, and 59.7% White patients; 29.5% admitted) and the BCH cohort included 22 222 visits (mean [SD] age, 8.2 [6.8] years; 52.1% male patients; 0.1% American Indian or Alaska Native patients, 4.0% Asian patients, 19.7% Black patients, 30.6% Hispanic or Latino patients, 0.2% Native Hawaiian or Pacific Islander patients, 37.7% White patients; 16.3% admitted). Among MIMIC-IV groups, intersectional debiasing was associated with a reduced subgroup calibration error from 0.083 to 0.065 (22.3%), while marginal fairness debiasing was associated with a reduced subgroup calibration error from 0.083 to 0.074 (11.3%; difference, 11.1%); among BCH groups, intersectional debiasing was associated with a reduced subgroup calibration error from 0.111 to 0.080 (28.3%), while marginal fairness debiasing was associated with a reduced subgroup calibration error from 0.111 to 0.086 (22.6%; difference, 5.7%). Among MIMIC-IV groups, intersectional debiasing was associated with lowered subgroup false-negative rates from 0.142 to 0.125 (11.9%), while marginal debiasing was associated with lowered subgroup false-negative rates from 0.142 to 0.132 (6.8%; difference, 5.1%). Fairness improvements did not decrease overall accuracy compared with baseline models (eg, MIMIC-IV: mean [SD] AUROC, 0.85 [0.00], both models). Intersectional debiasing was associated with lowered error rates in several intersectional subpopulations compared with other strategies. This study suggests that intersectional debiasing better mitigates performance disparities across intersecting groups than marginal debiasing for admission prediction. Intersectionally debiased models were associated with reduced group-specific errors without compromising overall accuracy. Clinical risk prediction models should consider incorporating intersectional debiasing into their development.

How Foundational Is the Retina Foundation Model? Estimating RETFound's Label Efficiency on Binary Classification of Normal versus Abnormal OCT Images.

How Foundational Is the Retina Foundation Model? Estimating RETFound's Label Efficiency on Binary Classification of Normal versus Abnormal OCT Images.

Development and validation of a deep learning algorithm for the classification of the level of surgical difficulty in impacted mandibular third molar surgery

Development and validation of a deep learning algorithm for the classification of the level of surgical difficulty in impacted mandibular third molar surgery

Optimized Faster R-CNN with Swintransformer for Robust Multi-Class Wildfire Detection

Wildfires are a critical global threat, emphasizing the need for efficient detection systems capable of identifying fires and distinguishing fire-related from non-fire events in their early stages. This study integrates the swintransformer into the Faster R-CNN backbone to overcome challenges in detecting small flames and smoke and distinguishing complex scenarios like fog/haze and chimney smoke. The proposed model was evaluated using a dataset comprising five classes: flames, smoke, clouds, fog/haze, and chimney smoke. Experimental results demonstrate that swintransformer-based models outperform ResNet-based Faster R-CNN models, achieving a maximum mAP50 of 0.841 with the swintransformer-based model. The model exhibited superior performance in detecting small and dynamic objects while reducing misclassification rates between similar classes, such as smoke and chimney smoke. Precision–recall analysis further validated the model’s robustness across diverse scenarios. However, slightly lower recall for specific classes and a lower FPS compared to ResNet models suggest a need for further optimization for real-time applications. This study highlights the swintransformer’s potential to enhance wildfire detection systems by addressing fire and non-fire events effectively. Future research will focus on optimizing its real-time performance and improving its recall for challenging scenarios, thereby contributing to the development of robust and reliable wildfire detection systems.