This doctoral research project aims to advance innovative diagnostic methodologies grounded in machine learning, using data collected through systematic field research across clinical and community contexts. The study identifies patterns and establishes associations between socio-historical living conditions and specific mental health pathologies — a line of investigation developed continuously over 15 years. The applied goal is the development of artificial intelligence software capable of formulating more effective prevention and intervention strategies, operationalized through biopsychosocial indicators.
The theoretical foundation integrates socio-historical psychology with computational epidemiology, treating subjectivity not as an isolated variable but as a dynamic product of structural conditions, biographical trajectory, and biological constitution. Machine learning is applied in the analysis of large datasets to identify patterns and associations between social factors, living conditions, and health outcomes that remain invisible to classical statistical approaches.
Structured interviews are conducted in field settings by trained research assistants, with data subsequently harmonized, semantically normalized, and projected into a standardized feature space for modeling. Variables span three analytical domains — biological (age of onset, chronicity, comorbidities), psychological (symptom profile, affect, cognition), and social (community networks, socioeconomic vulnerability, access barriers, adverse events). All variables are normalized to a continuous [0, 100] scale with effect direction preserved, enabling cross-domain comparison within the predictive model.
This analytical cross-sectional study develops and internally validates the Psychosocial Risk Index (IRP) — a composite instrument designed to translate the complexity of biopsychosocial determinants into a single operational risk score applicable in clinical triage and public health prioritization. The database comprises 1,155 structured interviews collected between October 2024 and January 2026 across 83 municipalities in 12 countries, spanning primary care units, psychosocial care centers, hospitals, and community social assistance services.
The outcome variable is operationalized at three levels: absence of diagnosis, general clinical diagnosis, and mental health diagnosis — used as a clinical risk proxy in multinomial logistic regression. Independent variables are derived from the structured interview, covering biological, psychological, and social dimensions. After semantic standardization, categorical variables are recoded as binary or ordinal features; continuous variables are normalized to [0, 100] with effect direction preserved.
The IRP organizes determinants into six operational domains: Structural, Traumatic, Stigma and Access, Internalizing, Functional, and Protective. Domain scores are computed as weighted composites of constituent variables, with weights derived from standardized multinomial regression coefficients and interaction terms capturing non-linear co-occurrence effects. Internal validation combines bootstrap resampling (1,000 iterations) with 5-fold cross-validation, evaluated by AUC-ROC, sensitivity, and specificity. The full analytical pipeline is implemented in Python 3.11 using pandas, numpy, statsmodels, scikit-learn, and matplotlib.
This study develops and validates a multimodal AI model capable of differentiating pseudoprogression from real tumor progression in high-grade gliomas — one of the most critical diagnostic challenges in neuro-oncology, where misclassification leads directly to inappropriate therapeutic decisions and unnecessary invasive procedures. Using data from approximately 900 patients (2015–2025), the model integrates multiparametric MRI radiomics with clinical-demographic variables and psychosocial factors, testing the hypothesis that multimodal fusion yields substantially higher diagnostic certainty than image-only approaches.
The study involves construction of a multimodal database, extraction and harmonization of radiomic features using advanced segmentation and standardized preprocessing following Radiomics Quality Score (RQS) guidelines, and application of deep learning architectures with early, intermediate, and late fusion strategies. The outcome is defined by specialist reviewers blinded to model outputs, ensuring unbiased ground truth labeling.
Model performance is evaluated using AUC-ROC, sensitivity, and specificity, with additional analyses of algorithmic fairness across demographic subgroups. Interpretability is ensured through explainable AI methods including Grad-CAM++ for spatial feature attribution in imaging data and SHAP values for clinical variable contribution. The study follows TRIPOD-ML reporting guidelines and contributes data from a Latin American population underrepresented in the international literature. Secondary objectives include simulating therapeutic decision scenarios to estimate clinical impact on reduction of unnecessary biopsies and inadequate treatment protocols.
The ability to understand and interpret human emotions is a critical dimension of interpersonal interaction and communication. Through advances in artificial intelligence — particularly in computer vision — machines can now play a significant role in detecting and analyzing human emotions from facial expressions and other visual signals. This study explores the application of computer vision to emotion detection from a psychological theory-informed perspective, aiming to bridge the technical and humanistic dimensions of affective computing in clinical mental health contexts.
The primary objective is to explore the viability and effectiveness of computer vision techniques for detecting human emotions, translating facial and visual expressions into structured numerical data through a Likert-scale quantification framework. The proposed interface processes visual input in real time, classifying emotional states into discrete and dimensional categories grounded in psychological theories of emotion — including basic emotion models and dimensional valence-arousal approaches.
The research adopts an interdisciplinary methodology integrating convolutional neural network (CNN) architectures trained on annotated facial expression datasets with psychological conceptual frameworks. Detected emotion vectors are mapped to Likert-scale output profiles, enabling their integration into structured clinical assessments and longitudinal monitoring tools. A central concern of the project is explainability: the system surfaces the visual features driving its classifications, supporting clinical trust and enabling practitioners to evaluate the model's reasoning — aligned with the broader commitment to transparent and auditable AI in healthcare.