Chemotherapy-induced peripheral neuropathy (CIPN) represents a frequent dose-limiting toxicity in oncology patients treated with neurotoxic agents such as taxanes and platinum compounds. Natural compounds with neuroprotective properties demonstrate supportive potential, particularly in preclinical studies, although their clinical integration remains limited due to insufficient practical tools and heterogeneous clinical evidence. A basic interactive prototype was developed in Python using Google Colab and the pandas and matplotlib libraries. The tool evaluates CIPN severity through a 0–4 scoring system for five common symptoms (numbness, tingling, pain, difficulty with fine motor tasks, and balance issues), achieving a maximum score of 20 (5 symptoms × 4 points). It computes the total score, provides an approximate grading according to NCI-CTCAE v5.0 criteria for Peripheral Sensory Neuropathy, and, in cases of elevated risk, displays a detailed table summarizing selected natural compounds (curcumin, quercetin, resveratrol, omega-3 fatty acids, vitamin E) including mechanisms of action, separate preclinical and clinical evidence levels, studied doses, precautions, and potential interactions. Results are presented in the form of a bar chart for symptom profile visualization. Preliminary testing on simulated patient profiles confirmed instant risk stratification and intuitive usability without installation requirements. This simple, accessible informatics prototype supports oncology clinicians in standardized CIPN monitoring and facilitates discussions with patients during consultations regarding adjacent natural supportive options while taking into consideration pharmacological safety. It offers a low-cost solution suitable for clinical extension in Romanian medical practice, such as mobile applications or electronic health record integration. AI tools assisted in code development and drafting; the prototype and final content represent original work by the author.
Background: The efficacy of medication in MDT is related to the number of loaded magnetic particles at the target site. Accurate quantification of this deposited dose is important for assessing treatment effectiveness and safety, and for enabling patient-specific MDT planning and in silico modelling to emulate in vivo efficacy. Particle deposition is transient under vascular flow: particles first accumulate rapidly, then partially erode. Static or hand-selected frame acquisition does not account for these dynamics, reducing experimental realism and challenging the generation of patient-specific data. Objective: To establish an automated, video-based experimental workflow that allows depiction of peak particle deposition and deposited drug dose with high time resolution and accuracy by generating structured datasets for virtual dose estimation, deep learning, and patient-specific MDT planning within digital twin concepts. Methods: The integrated medical-informatics pipeline includes a high-resolution multi-angle acquisition system with semi-automatic 3-axis magnetic positioning and synchronised views of a biologically relevant microvasculature structure. Time series of video streams are analysed using computer-vision techniques to segment deposition sites, extract quantitative image-derived data, and inform time-dependent deposition dynamics. These are combined to generate temporal deposition profiles, from which peak deposition is automatically detected. Areas from orthogonal projections of the deposits are integrated under the assumption of a specific geometric reconstruction to estimate 3D deposition volume, and mass (and, by inference, delivered dose) is obtained using calibration data. Annotated video frames, peak-deposition labels, magnet-position parameters, and temporal metadata are converted into structured learning-ready datasets for supervised deep learning and surrogate modelling. Results: The platform enables observer-independent, time-dependent estimation of the deposited drug amount and surrogate dose determination under dynamically flowing conditions. Conclusions: The proposed platform supports patient-specific MDT planning, integration into digital twin models, and data-driven optimisation of magnetic drug targeting strategies by converting experimental video data into structured datasets.
Background: Medical scores represent simplified methods for estimating the prognosis, risk or severity of a condition and are also used in therapeutic decisions, triage and resource allocation, assessment of functionality and frailty, standardization of medical communication. Medical scores add support in the assessment of health status, providing a common, objective and comparable framework for decision-making, but do not replace clinical judgment. Applied medicine must be objective and must eliminate the subjective form caused by the human factor. This is achieved by transforming clinical data into evidence-based decisions - thus reducing human errors and ensuring that all patients benefit from an equally optimized and efficient treatment. Methods: Defining score is a process that follows a rigorous statistical methodology to ensure that the score is predictive, reproducible, and applicable in practice. The steps required to define the score: • Data collection, • Definition of variables for inclusion, • Model design and optimization (Linear, Logistic, Cox), • Defining the score, • Score performance (Se, Sp, ROC curve), • Validation. New methods for scoring (inspired by Machine Learning): Fuzzy Systems, Neural Networks, Classifications, Clustering, Dimensions Reduction (Principal Components). Given the power of computers, complex scoring techniques can be applied, but the influence of covariates and factors would be more difficult or even impossible to quantify (example: neural networks). The scores are not perfect and as a result have disadvantages: they do not capture all individual peculiarities, they can be applied rigidly, and they sometimes underestimate or overestimate risk.One method of compensating for these errors is to adapt the scores to the target population. Conclusions: Scores represent an important medical support, providing information through a simplified form of calculation by quantifying the effect of covariates. The bridge between the complex reality of medical cases and statistical-mathematical rigor is made by scores in a simplified but useful form.
Systematically extracting adverse event (AE) data from Summary of Product Characteristics (SmPC) documents is often necessary in exploring drug safety with computational means (e.g. for building QSAR models). The European Medicines Agency (EMA) makes available the SmPCs in a pdf format, which complicates this process. Inconsistent formatting (at least two different ways of structuring tables), multi-page tables, variable number of tables included in section 4.8, and varied terminology, combined with ambiguous frequency data, create significant barriers to reliable, large-scale (semi)automated analysis. We built a two‑stage extraction workflow in R that integrates table‑based and text‑based methods. Using the “tabulapdf” package, we first retrieved AE tables spanning multiple pages, automatically removed repeated headers, and merged fragmented tables. We then applied a rule‑driven text parser focused on 10 monoclonal‑antibody‑related AEs (anaphylaxis, cytokine release syndrome, angioedema, urticaria, rash, pyrexia, hypersensitivity, bronchospasm, hypotension, pruritus). The parser handles terminology variability through synonym mapping, prioritizes longer strings to avoid partial matches, and uses context‑sensitive frequency detection at both line-level and section-level frequency indicators. Frequencies are mapped to a standardized 0-6 scale (0=absent, 1=unknown frequency, 2-6=very rare to very common). Outputs from the extraction tool are standardized, providing detailed metrics: frequency scores for each AE, aggregated counts of all AEs and serious cases (including anaphylaxis, cytokine release syndrome, and angioedema), and the peak frequency among reported AEs. The method is robust in processing nested tables, preserving the connection between frequencies and AEs despite line breaks, and correctly distinguishes between AEs that are absent from those with unreported frequencies. Performance validation challenges are addressed, alongside comparisons with available open-source and proprietary tools. The pipeline allows consistent and reproducible extraction of adverse event information from EMA SmPCs, supporting scalable comparative safety and pharmacovigilance analyses.
Kardio Diagnostix was founded with the goal of reducing the burden on health systems while increasing the quality of care. We have developed a a novel deep learning architecture integrating a physics-informed Harmonic-Percussive-Residual (HPR) signal decomposition with a stacked Long Short-Term Memory (LSTM) network to screen heart murmurs found in 3.2 million children yearly in North America alone. Our platform has the potential to reduce the number of unnecessary referrals of innocent murmurs to pediatric cardiologists in North America from 1.6 million to 75,000, reducing the cost of care by US$3.8 billion annually. Besides cost reduction, we will increase the detection of abnormal murmurs by 75%. Our product has a potential to access an annual US$1.5 billion market for the care of children with murmurs in North America. We tested our product with heart murmurs for 300 patients in Nova Scotia. The data was collected he data was obtained by the cardiologists or delegated research assistant for the patients who arrived at their clinic during the study period. We are expanding our data collection to other regions starting with a clinic in Toronto. We describe a comprehensive, data science-driven methodology for developing and deploying any intelligent medical diagnostics application. This strategy centers on a cyclical process that ensures the diagnostic model is accurate, integrated, compliant, and continuously improving within a clinical environment. The foundational stage involves establishing a secure, scalable architecture for handling sensitive medical data, including implementing a real-time data ingestion and feedback integration engine and designing data pipelines for EMR data synchronization. Next, the diagnostic model must be validated at scale using diverse, real-world clinical data to ensure accuracy. Concurrently, co-design sessions and field trials help develop production-ready UI/UX that aligns with clinical workflows. Crucial for long-term use is the MLOps automation phase, which involves setting up CI/CD pipelines, real-time monitoring, and auto-retraining mechanisms to establish a continuous learning loop. Finally, to ensure deployment readiness, the system must integrate robust features for secure data handling, access controls, and audit logging to meet compliance standards (e.g., HIPAA/PIPEDA). Implementing Explainable AI (XAI) dashboards provides necessary transparency for clinicians, supporting the final deployment.
