Artificial intelligence (AI) uses mathematical algorithms to perform tasks that require human cognitive abilities. AI-based methodologies, e.g., machine learning and deep learning, as well as the recently developed research field of radiomics have noticeable potential to transform medical diagnostics. AI-based techniques applied to medical imaging allow to detect biological abnormalities, to diagnostic neoplasms or to predict the response to treatment. Nonetheless, the diagnostic accuracy of these methods is still a matter of debate. In this article, we first illustrate the key concepts and workflow characteristics of machine learning, deep learning and radiomics. We outline considerations regarding data input requirements, differences among these methodologies and their limitations. Subsequently, a concise overview is presented regarding the application of AI methods to the evaluation of thyroid images. We developed a critical discussion concerning limits and open challenges that should be addressed before the translation of AI techniques to the broad clinical use. Clarification of the pitfalls of AI-based techniques results crucial in order to ensure the optimal application for each patient.

Although rare, immune checkpoint inhibitor (ICI)-related myocarditis can be life-threatening, even fatal. In view of increased ICI prescription, identification of clinical risk factors for ICI-related myocarditis is of primary importance. This study aimed to assess whether pre-existing cardiovascular (CV) patient conditions are associated with the reporting of ICI-related myocarditis in VigiBase, the WHO global database of suspected adverse drug reactions (ADRs). In a (retrospective) matched case-control study, 108 cases of ICI-related myocarditis and 108 controls of ICI-related ADRs other than myocarditis were selected from VigiBase. Drugs labeled as treatment for CV conditions (used as a proxy for concomitant CV risk factors and/or CV diseases) were found to be associated more strongly with the reporting of ICI-related myocarditis than with other ICI-related ADRs (McNemar’s chi-square test of marginal homogeneity: p = 0.026, Cramer’s coefficient of effect size: Φ = 0.214). No significant association was found between pre-existing diabetes and ICI-related myocarditis reporting (McNemar’s test of marginal homogeneity: p = 0.752). These findings offer an invitation for future prospective pharmacoepidemiological studies to assess the causal relationship between pre-existing CV conditions and myocarditis onset in a cohort of cancer patients followed during ICI treatment.

Background: Thyroid nodule (TN) is a common entity, and TNs assessed by Thyroid Imaging and Reporting Data Systems (TIRADSs) as low-risk lesions (TIRADS 3) can significantly impact the rate of unnecessary biopsy (UN-FNAC). This study reviewed a consecutive series of patients undergoing surgery to analyze TNs assessed as TIRADS 3. Methods: Thyroid surgeries performed from January 2019 to August 2024 were reviewed. Patients with preoperative thyroid ultrasound were selected, and TNs were classified according to American College of Radiology (ACR), European Thyroid Association (EU), and Korean (K) TIRADS. Cases assessed as TIRADS 3 were finally included. Histology was the reference standard to calculate the rate of UN-FNAC. Results: The study series included 284 TNs assessed as TIRADS 3. The risk of malignancy was 8.7% in ACR-, 10.7% in EU-, and 10.1% in K-TIRADS, higher than expected. The frequency of TNs with indication for biopsy according to K-TIRADS (66.7%) was significantly (P = 0.003) higher than ACR-TIRADS (46.7%), with an intermediate value of EU-TIRADS (56.5%). The percentage of cancers with indication for biopsy according to ACR-, EU-, and K-TIRADS was 25%, 50%, and 50%, respectively. The overall rate of UN-FNAC was 95.3% in ACR-TIRADS, 90.3% in EU-TIRADS, and 92.4% in K-TIRADS. Conclusion: How to save on UN-FNACs in low-risk TNs is challenging. Although ACR-TIRADS can be effective in reducing the total number of biopsies, the rate of UN-FNAC remains significant. Alternative strategies should be developed.

Background: Acute kidney injury (AKI) and acute kidney disease (AKD) are frequent complications of hospitalization, resulting in reduced outcomes and increased cost burden. However, these conditions are only sometimes recognized and promptly treated. Leveraging electronic health records (EHR), we explored the potential of artificial intelligence (AI) in diagnosing AKI and AKD during hospitalization. Methods: We retrospectively analyzed EHRs collected from all patients admitted in 2022 to our public hospital network. AKI and AKD were defined according to international guidelines. The database was divided into training and validation sets. Machine Learning (ML) algorithms were developed with 10-fold cross-validation, and diagnostic accuracy was evaluated. Findings: We analyzed 34,579 hospitalizations (mean age of 60 years, 50% females). Baseline renal function was available in ~50% of cases. AKI and AKD complicated 10% and 1.5% of hospitalizations, respectively. The majority of AKI episodes (77%) occurred within the first three days of hospitalization, and >50% of subjects with AKI were discharged before complete renal function recovery. ML accurately predicted AKI (AUC-ROC 79%) during hospitalization, based on data available before and at hospital admission. Among subjects with AKI on the first day and longer in-hospital observation, the ML accuracy in predicting AKD transition increased (AUC-ROC from 76% to 88%) by integrating EHR accumulated during the hospitalization. The negative predictive value (NPV) progressively increased from 94% to 98% consistently. Shapely additive explanations documented that age, urgent hospital admission, AKI severity, and baseline renal function were associated with AKI. Renal function trajectory during the first days of hospitalization was the most relevant predictor of AKD. Interpretation: ML, relying on EHR before and during hospitalization, may accurately predict AKI and AKD. The high NPV also suggests its implementation as a tool to rule out the risk of renal failure, aid in individualizing patient care, and allocate healthcare resources.

Objectives: Ultrasound (US) evaluation is recognized as pivotal in assessing the risk of malignancy (RoM) of thyroid nodules (TNs). Recently, various US-based risk-classification systems (Thyroid Imaging and Reporting Data Systems [TIRADSs] have been developed. An important ongoing project concerns the creation of an international system (I-TIRADS) using unique terminology. Since online tool allow clinicians and patients to stratify the RoM of any TN, the role of computer scientist (CS) should be relevant. This study explored the performance of CS in assessing TNs across the TIRADS categories. Methods: The most diffused TIRADSs (i.e., ACR, EU, and K) were considered. Three-hundred scenarios were created. A CS was asked to assess the 300 TNs according to ACR-, EU-, and K-TIRADS. These data were compared with that of clinicians. The inter-observer agreement was estimated with Cohen kappa (κ). Word-cloud plots were used to graph the US descriptors with disagreement. Results: The correspondence of the CS’s assessment with the physicians was 100%, 81%, and 43%, using ACR-, EU-, and K-TIRADS, respectively. The CS was unable to classify 19/100 TNs according to EU-TIRADS and 15/100 TNs according to K-TIRADS. The inter-observer agreement between CS and physicians was excellent for ACR-TIRADS (κ = 1), moderate for EU-TIRADS (κ = 0.56), and fair for K-TIRADS (κ = 0.22). Among the non-concordant cases, 16/22 descriptors for EU-TIRADS and 18/18 descriptors for K-TIRADS were found. Conclusion: CSs are confident with the ACR-TIRADS lexicon and structure while not with EU- and K-TIRADS, probably because they are pattern-based systems requiring medical training.