Indoor monitoring of biological contamination is essential for protecting cultural heritage and public health. However, conventional culture-based methods limit timely intervention. This study presents an affordable modular multisensor system for indirectly detecting airborne fungal contamination using Penicillium chrysogenum as a representative model organism and its environmental signatures. The proposed prototype integrates PMSA003I, BME688 and AMG8833 sensors and was evaluated under controlled environmental conditions. Biological ground truth was established using a volumetric inertial-impaction sampling protocol (SAS sampler), validating four contamination levels (~6 to 165, CFU/m³). A total of 1989 observations were analyzed. Non-parametric statistical tests (Kruskal–Wallis and Mann–Whitney U) confirmed significant differences between all the exposure conditions ($p<0.001$). Supervised machine learning (ML) models showed strong performance across all the classification tasks, with accuracy and AUC values near 100%. In most cases, pressure alone was sufficient. The statistical and ML analyses consistently identified pressure, particulate-related variables, gas resistance and humidity as the most informative features. Overall, the results indicate that the proposed approach can reliably capture indirect environmental signatures associated with airborne fungal presence under controlled conditions. The study supports the feasibility of low-cost multisensor systems for continuous indoor bioaerosol monitoring while highlighting the need for further optimization and validation in real-world environments.

Alzheimer’s disease (AD) is the leading cause of dementia and a major cause of death worldwide, making early detection a critical clinical priority. Because pathological changes may begin 15–20 years before symptom onset, artificial intelligence (AI) has emerged as a promising tool for identifying and characterizing AD. In particular, multi-modal approaches that integrate cognitive, biological, and sensor-based data have attracted growing interest. This systematic review compares single- and multi-modal AI strategies for AD detection, covering machine learning and deep learning methods, feature representations, validation strategies, and classification tasks. Searches of major databases identified 568 studies published between 2016 and early 2026; 278 met the inclusion criteria according to PRISMA guidelines. Multi-modal approaches generally achieved higher performance than single-modal strategies, particularly for challenging tasks such as predicting progression between closely related disease stages, although direct comparisons under identical conditions remain scarce. Critically, only about 4% of studies evaluated their models on a genuinely independent external cohort, raising substantial concerns about model generalizability. Overall, current AI systems remain highly dependent on existing datasets and heterogeneous evaluation protocols, which limit generalizability and clinical applicability. Future research should prioritize representative multimodal datasets, rigorous external validation, and clinically interpretable AI systems.