Sandeep Prakash

Pica is a clinically significant and often overlooked feeding and eating disorder in which individuals have a persistent urge to eat or chew non-nutritious or non-food items and is linked to nutritional, hematological, behavioral, and environmental issues. Human studies have regularly found links between pica and iron deficiency and anemia, as well as between changes in hematological parameters, decreased ferritin, and decreased zinc concentrations; toxic-element exposure may also be a part of certain behaviors (like geophagia). The review critically summarises the potential use of multimodal artificial intelligence (AI) and machine learning (ML) to support predictive risk stratification, early detection, and clinical management of pica, highlighting the importance of combining etiological and pathophysiological biomarkers, clinical, behavioural, toxicological and electronic health-record data. Biomarker evidence also indicates human exposure to multiple biomarkers together may be a more complete risk profile than single biomarkers, since there are human data from which to draw conclusions. AI and natural language processing could be used to identify undetected pica-related behaviors and track changes over time to clinical and laboratory data. However, there are still limitations in the development and validation of pica-specific AI models, and existing evidence from related conditions mainly methodologically supports the use of AI. Further studies are needed on large, prospective, diverse human cohorts, standardized pica phenotyping, longitudinal assessment of biomarkers, explainable AI, external validation, and prospective clinical assessment. Finally, multimodal AI should be used as a clinically interpretable decision support system, not as an independent diagnostic system.

These rare facial diseases, which include hemifacial microsomia, Parry–Romberg syndrome, Moebius syndrome, Treacher Collins syndrome, Apert syndrome, and Crouzon syndrome, among others, pose considerable difficulties in diagnosis and treatment due to their clinical variability, rare occurrence, and complex genetic etiology. This review highlights the increasing potential of artificial intelligence (AI) and machine learning (ML) technologies in the better diagnosis, phenotyping, genotyping, and clinical management of these rare disorders. In this review, the applications of AI to facial phenotyping, three-dimensional (3D) imaging, radiomics, multimodal learning, and explainable AI have been highlighted. It also points out some latest developments in the field of bioinformatics, genome editing, RNA therapies, patient-derived models, and digital twin technology for precision medicine and translational research. Presently, there is sufficient evidence indicating the benefits of AI in increasing the accuracy in diagnosis, objective craniofacial evaluation, and customized treatment plans, whereas computational therapy is still experimental. However, various barriers such as data limitations, variability in phenotype, bias in the algorithms, external validation, ethical issues, and regulation prevent its widespread adoption in clinical practice. Future studies need to concentrate on multicenter data sharing, multimodal explainable AI, precision genomics, and translational framework to speed up the adoption of AI in rare craniofacial medicine.

Rare appendix malignancies, such as mucinous adenocarcinoma, goblet cell adenocarcinoma, signet-ring cell carcinoma, and appendiceal neuroendocrine tumors, are rare gastrointestinal cancers with diagnostic and treatment challenges due to their histological and molecular diversity. This study presents an explainable Artificial Intelligence (AI)-based approach that integrates digital histological biopsy images, molecular markers, and patient data to enhance disease classification and prediction of response to treatment. A quantitative methodology was used, based on an anonymized dataset of 200 patients with rare appendix malignancies. The performance of machine learning algorithms, such as Logistic Regression, Random Forest, XGBoost, and LightGBM, was compared using Accuracy, Precision, Recall, F1 score, and ROC-AUC metrics, while SHAP and Grad-CAM methods increased model interpretability. LightGBM outperformed all other methods and achieved the highest accuracy (96.2%), F1-score (95.5%), and ROC-AUC (0.987). KRAS, Ki-67, and TP53 were determined to be the most significant predictors from the biomarker analysis and the overall treatment response prediction accuracy of the combined model was found to be 95.1%. These results show that combining digital histopathology, biomarkers, and explainable AI (XAI) can enhance diagnosis accuracy, treatment response predictions, and clinical decision-making in rare appendiceal malignancies.