Daily Endocrinology Research Analysis
Three impactful endocrinology-adjacent studies stand out today: an explainable multimodal AI robustly predicts lateral lymph node metastasis preoperatively in thyroid cancer; SGLT2 inhibitor initiation is associated with lower incident dementia in patients with heart failure and diabetes; and 18F-choline PET enables accurate localization to support focused parathyroidectomy after inconclusive conventional imaging. These findings span diagnostic innovation, pharmacoepidemiology, and surgical plan
Summary
Three impactful endocrinology-adjacent studies stand out today: an explainable multimodal AI robustly predicts lateral lymph node metastasis preoperatively in thyroid cancer; SGLT2 inhibitor initiation is associated with lower incident dementia in patients with heart failure and diabetes; and 18F-choline PET enables accurate localization to support focused parathyroidectomy after inconclusive conventional imaging. These findings span diagnostic innovation, pharmacoepidemiology, and surgical planning.
Research Themes
- Explainable AI for endocrine oncology risk stratification
- Cardiorenal-metabolic therapeutics and neurocognitive outcomes
- Advanced imaging to enable minimally invasive endocrine surgery
Selected Articles
1. Explainable multimodal deep learning for predicting thyroid cancer lateral lymph node metastasis using ultrasound imaging.
Using 29,615 multimodal preoperative cases, LLNM-Net accurately predicted lateral nodal metastasis in thyroid cancer, outperforming radiologists and prior models. Capsule-adjacent tumors (<0.25 cm) and upper/middle lobe locations were identified as high-risk features, enabling fine-grained preoperative risk stratification.
Impact: This study demonstrates explainable, multicenter-validated AI that materially outperforms experts for a decision-critical task in thyroid surgery planning.
Clinical Implications: Supports selective lateral neck dissection by identifying high-risk patients preoperatively, potentially reducing unnecessary dissections and complications.
Key Findings
- LLNM-Net achieved AUC 0.944 and 84.7% accuracy in multicenter testing, outperforming human experts (64.3%).
- Model surpassed prior approaches by 7.4% absolute accuracy.
- Tumors within 0.25 cm of the thyroid capsule had >72% metastasis risk; upper/middle lobe tumors were high-risk regions.
- High-risk identification achieved AUC 0.983 by integrating imaging features with clinical text and demographics.
Methodological Strengths
- Large, multicenter dataset with external testing across seven centers
- Multimodal fusion (ultrasound images, reports, pathology, demographics) with explainability
- Benchmarking against human experts and prior models
Limitations
- Retrospective design with potential selection and information bias
- Generalizability to other healthcare systems and devices beyond participating centers remains to be proven
- Clinical impact not yet tested in prospective interventional trials
Future Directions: Prospective, multicenter clinical trials to evaluate decision impact and patient outcomes; integration into workflow with clinician-in-the-loop validation; assessment across diverse ultrasound vendors and populations.
Preoperative prediction of lateral lymph node metastasis is clinically crucial for guiding surgical strategy and prognosis assessment, yet precise prediction methods are lacking. We therefore develop Lateral Lymph Node Metastasis Network (LLNM-Net), a bidirectional-attention deep-learning model that fuses multimodal data (preoperative ultrasound images, radiology reports, pathological findings, and demographics) from 29,615 patients and 9836 surgical cases across seven centers. Integrating nodule morphology and position with clinical text, LLNM-Net achieves an Area Under the Curve (AUC) of 0.944 and 84.7% accuracy in multicenter testing, outperforming human experts (64.3% accuracy) and surpassing previous models by 7.4%. Here we show tumors within 0.25 cm of the thyroid capsule carry >72% metastasis risk, with middle and upper lobes as high-risk regions. Leveraging location, shape, echogenicity, margins, demographics, and clinician inputs, LLNM-Net further attains an AUC of 0.983 for identifying high-risk patients. The model is thus a promising for tool for preoperative screening and risk stratification.
2. Sodium-glucose Cotransporter-2 Inhibitor Initiation and Incident Dementia in Heart Failure With Diabetes: A Population-based Cohort Study.
In a target trial emulation using Ontario administrative data, initiating an SGLT2 inhibitor versus a DPP4 inhibitor was associated with lower incident dementia among older adults with heart failure and diabetes over a median 3.95 years. Effect estimates were robust across ITT and as-treated analyses using propensity weighting and competing risk methods.
Impact: Connects a cornerstone cardiorenal therapy to a major neurocognitive outcome in a high-risk population using rigorous causal inference methods.
Clinical Implications: When selecting glucose-lowering therapy in heart failure with diabetes, SGLT2 inhibitors may offer neurocognitive benefits in addition to established cardiorenal protection.
Key Findings
- Initiation of SGLT2 inhibitors vs DPP4 inhibitors was associated with lower incident dementia (HR 0.73; 95% CI 0.60-0.87; IRD -8.1/1000 person-years).
- As-treated analysis suggested larger effect size (HR 0.53; 95% CI 0.39-0.70; IRD -14.2/1000 person-years).
- Design used a 180-day exposure lag, propensity-score fine stratification, and competing risk models.
Methodological Strengths
- Target trial emulation with intention-to-treat and as-treated analyses
- Propensity-score fine stratification weights and competing risk modeling
- New-user active-comparator design using population-based data
Limitations
- Observational design susceptible to residual confounding and misclassification
- Use of administrative coding may limit phenotypic granularity and dementia subtype classification
- Medication adherence and crossover cannot be fully captured
Future Directions: Prospective pragmatic trials or advanced emulations with richer cognitive phenotyping to confirm causality; mechanistic studies on SGLT2i neuroprotection in HF.
BACKGROUND: Heart failure often coexists with important dementia-risk factors, such as diabetes, atrial fibrillation and hypertension. Sodium-glucose cotransporter-2 (SGLT2) inhibitors have been associated with a lower dementia risk in general diabetes populations, but evidence is limited, specifically in heart failure with comorbid diabetes. OBJECTIVE: To investigate the association of SGLT2 inhibitors with incident dementia in people with heart failure and diabetes. METHODS: This target trial emulation cohort study used linkable administrative databases from Ontario, Canada. New users of SGLT2 inhibitors or dipeptidyl peptidase-4 (DPP4) inhibitors aged ≥ 66 years with diabetes and heart failure (July 2016-December 2020) entered this cohort. A 180-day lag time was implemented to mitigate reverse causality. The primary analysis used an intention-to-treat exposure definition. Cause-specific hazard ratios (HRs), with death as a competing risk, were estimated by using Cox models with propensity-score fine stratification weights. Weighted incidence-rate differences (IRDs) per 1000 person-years were also estimated. RESULTS: Among 4402 SGLT2 inhibitor and 6319 DPP4 inhibitor new users, over a median follow-up of 3.95 years from treatment initiation, SGLT2 inhibitor vs DPP4 inhibitor initiation was associated with lower dementia risk (HR 0.73, 95% confidence interval [CI] 0.60-0.87; IRD -8.1, 95% CI -12.7 to -3.5). The secondary as-treated analysis showed greater risk reduction (HR 0.53, 95% CI 0.39-0.70; IRD -14.2, 95% CI -20.1 to -8.4) than the primary intention-to-treat analysis. CONCLUSIONS: SGLT2 inhibitor initiation was associated with dementia risk reduction in heart failure and diabetes, a population at a high risk of developing dementia.
3. Impact of 18F-choline PET-CT or PET-MRI on surgical strategy in patients with primary hyperparathyroidism.
In 185 PHPT surgeries, adding 18F-choline PET after negative/inconclusive US/MIBI markedly improved localization sensitivity (94.4% vs 63.9%) and enabled unilateral focused approaches in most cases, with an overall 100% cure rate. FCh particularly supported focused parathyroidectomy in milder disease with smaller glands.
Impact: Demonstrates a practical imaging strategy that shifts surgery from bilateral exploration to focused approaches in one-third of difficult-localization PHPT cases.
Clinical Implications: Consider 18F-choline PET when US/MIBI are negative or inconclusive to facilitate focused parathyroidectomy and reduce operative morbidity.
Key Findings
- Imaging sensitivity improved from 63.9% (MIBI group) to 94.4% (FCh group; P < 0.001).
- FCh enabled clear unilateral localization in 86.9% of cases, avoiding unnecessary bilateral neck exploration in ~33%.
- Overall surgical cure rate was 100% across 185 PHPT patients.
- FCh group had lower pre-op calcium and PTH and smaller/lighter glands, suggesting utility in milder PHPT.
Methodological Strengths
- Direct comparison of imaging strategies with surgical outcomes across a full operative cohort
- Clinically relevant endpoints (localization enabling focused surgery, cure rate)
Limitations
- Retrospective, non-randomized comparison between groups
- Potential selection bias based on prior imaging results and referral patterns
- Single-system experience may limit generalizability
Future Directions: Prospective cost-effectiveness analyses and randomized pathways comparing FCh-first vs conventional imaging strategies; evaluation in normocalcemic/mild PHPT and multigland disease.
BACKGROUND: Accurate preoperative localization is essential for successful, focused, minimally invasive surgery in primary hyperparathyroidism (PHPT). New imaging techniques have recently been proposed. This study evaluated the impact of 18F-choline positron emission tomography (PET)-computed tomography or 18F-choline PET-magnetic resonance imaging (FCh) in patients with negative or inconclusive results on neck ultrasonography (US) and 99mTc-sestamibi (MIBI) scintigraphy. METHODS: Baseline biochemical characteristics (preoperative calcemia and PTH), parathyroid gland features (size and weight), preoperative imaging localization techniques accuracy, and surgical results were compared in a series of patients operated for PHPT who underwent only preoperative US and MIBI scintigraphy with concordant results (MIBI Group) or also FCh as additional imaging following US and MIBI with negative or inconclusive results (FCh Group). RESULTS: The overall cure rate was 100% in 185 patients operated for PHPT. The overall sensitivity of imaging was 63.9% in the MIBI group (n = 116), compared with 94.4% (P < 0.001) in the FCh group (n = 69). FCh provided clear unilateral localization in 86.9% of patients, avoiding unnecessary bilateral neck exploration; in contrast, based on MIBI results, unilateral localization would have been theoretically possible in only 61.6% of patients. Compared with the MIBI group, patients in the FCh group had significantly lower preoperative calcium levels (2.71 versus 2.79 mmol/l; P = 0.012), lower preoperative parathyroid hormone levels (177 versus 250 pg/ml; P = 0.032), and smaller (17 versus 21 mm; P <0.001) and lighter (1.47 versus 2.58 g, P = 0.005) parathyroid glands removed. CONCLUSION: FCh enables successful focused parathyroidectomy in PHPT patients with negative or inconclusive MIBI results, reducing unnecessary bilateral neck exploration in 33% of patients; it may also allow for a successful focused approach in patients with milder PHPT, characterized by lower preoperative calcium and PTH levels and smaller pathological parathyroid glands.