The aging of the world’s population, the willingness of elderly to remain independent, and the recent COVID-19 pandemic have demonstrated the urgent need for home-based diagnostic and patient monitoring systems to reduce the financial and organizational burdens that impact healthcare organizations and professionals. The Internet of Medical Things (IoMT), i.e., all medical devices and applications that connect to health information systems through online computer networks. The IoMT is one of the domains of IoT where real-time processing of data and reliability are crucial. In this paper, we propose RAMi, which is a Real-Time Architecture for the Monitoring of elderly patients thanks to the Internet of Medical Things. This new architecture includes a Things layer where data are retrieved from sensors or smartphone, a Fog layer built on a smart gateway, Mobile Edge Computing (MEC), a cloud component, blockchain, and Artificial Intelligence (AI) to address the specific problems of IoMT. Data are processed at Fog level, MEC or cloud in function of the workload, resource requirements, and the level of confidentiality. A local blockchain allows workload orchestration between Fog, MEC, and Cloud while a global blockchain secures exchanges and data sharing by means of smart contracts. Our architecture allows to follow elderly persons and patients during and after their hospitalization. In addition, our architecture allows the use of federated learning to train AI algorithms while respecting privacy and data confidentiality. AI is also used to detect patterns of intrusion.

The huge amount of data produced by the Internet of Things need to be validated and curated to be prepared for the selection of relevant data in order to prototype models, train them, and serve the model. On the other side, blockchains and open data are also important data sources that need to be integrated into the proposed integrative models. It is difficult to find a sufficiently versatile and agnostic architecture based on the main machine learning frameworks that facilitate model development and allow continuous training to continuously improve them from the data streams. The paper describes the conceptualization, implementation, and testing of a new architecture that proposes a use case agnostic processing chain. The proposed architecture is mainly built around the Apache Submarine, an unified Machine Learning platform that facilitates the training and deployment of algorithms. Here, Internet of Things data are collected and formatted at the edge level. They are then processed and validated at the fog level. On the other hand, open data and blockchain data via Blockchain Access Layer are directly processed at the cloud level. Finally, the data are preprocessed to feed scalable machine learning algorithms.

Parkinson's disease (PD) exhibits substantial heterogeneity in clinical presentation and longitudinal progression, complicating prognosis, trial design, and precision therapeutics. While the Subtype and Stage Inference (SuStaIn) algorithm family offers probabilistic disease progression modeling, comprehensive comparative analyses and real-world validation remain limited. To systematically compare major SuStaIn variants for PD progression modeling and validate their stability and prognostic utility using both controlled simulated data and real-world Parkinson's Progression Markers Initiative (PPMI) cohorts. We constructed a biologically grounded simulated PD cohort (n=400; 5 visits over 4 years; 12 motor and non-motor biomarkers) with three predefined progression subtypes reflecting realistic noise, missingness (7.7%), and effect sizes informed by PPMI literature. Six SuStaIn variants were benchmarked: Z-score, Ordinal, Event-based Mixture, Missing Data, s-SuStaIn, and Temporal SuStaIn. We then applied identical variants to baseline PPMI data (N=624; PD=427; prodromal=197) using MDS-UPDRS I-IV, MoCA, GDS, STAI, SCOPA-AUT, RBD screening, ESS, and UPSIT. We assessed cross-sectional subtype structure, test-retest stability (Cohen's κ), and prognostic performance for motor worsening, cognitive decline, dyskinesia, and advanced therapy using Cox models with concordance indices and 95% confidence intervals. Simulation analyses revealed six clinically interpretable biomarker domains (motor complications, olfactory, cognitive-sleep, neuropsychiatric/autonomic, early non-motor, motor-depression) and three progression subtypes (slow 26.5%, intermediate 46.8%, fast 26.8%). Temporal and s-SuStaIn variants achieved optimal balance of accuracy (ARI: 0.058 and 0.047), stability, and cluster separation (silhouette: 0.284 and 0.267). In PPMI validation, all variants converged on three subtypes: benign motor-predominant (30.0%, age 61.4 ± 9.2 years), intermediate mixed (42.9%, age 64.8 ± 8.7 years), and diffuse-malignant (27.1%, age 67.3 ± 7.9 years, PIGD-dominant). Augmenting clinical covariates with ensemble SuStaIn subtypes improved motor progression prediction (c-index: 0.674 [0.648-0.701] to 0.741 [0.718-0.765], Δc=+0.067, p<0.001) and cognitive decline prediction (0.628 [0.598-0.658] to 0.764 [0.739-0.789], Δc=+0.136, p<0.001). Ensemble fusion enhanced test-retest stability (Cohen's κ=0.81 [0.77-0.86]) and outperformed single variants by 15% (relative gain). A reduced panel (UPDRS I-III, MoCA, GDS, ESS, UPSIT; 7 biomarkers) retained 94% of full-model performance. SuStaIn variants capture complementary PD progression patterns. A principled ensemble strategy, validated on 624 PPMI participants, yields stable and prognostically informative subtypes. The diffuse-malignant subtype enables trial enrichment with 63% sample size reduction. Our dual-cohort framework establishes which algorithmic findings generalize to real-world constraints, providing a practical template for PD precision medicine.

Optimizing healthcare resources in neurodegenerative diseases requires balancing diagnostic performance with cost constraints. We introduce AHN-BudgetNet—a tiered, cost-aware assessment framework for Parkinson’s disease motor severity prediction—evaluated on 1387 simulated PPMI subjects via patient-level GroupKFold validation. Our analysis tested seven tier combinations encompassing demographic, self-reported, and clinical features. The baseline (T0) yields AUC = 0.65 (95% CI [0.629, 0.681]) at no cost. Self-assessments (T1) alone achieved an AUC = 0.69 (95% CI [0.643, 0.733]) at USD 75, with an efficiency of 1.07. The combined T0 + T1 set reached AUC = 0.75 (95% CI [0.729, 0.772]) at USD 75, with efficiency 1.43. T2 alone obtained AUC = 0.53 (95% CI [0.517, 0.542]) at USD 300 and efficiency 0.07. The full T0 + T1 + T2 set achieved the highest performance—AUC = 0.76 (95% CI [0.735, 0.774])—at USD 375, with efficiency 0.54, reflecting diminishing returns beyond T1. High-cost tiers (T3/T4) could not be empirically validated due to over 88% missing data, emphasizing the value of accessible assessments. Gaussian Mixture on Tier 0 features yielded a silhouette score of 0.54, compared to 0.53 for K-means, confirming that patient-reported outcomes can support clinical stratification. Our results underpin evidence-based resource allocation: budgets USD ≤ 75 prioritize T1, while budgets USD ≤ 375 justify a comprehensive assessment. This confirms that structured tier prioritization supports robust, resource-efficient diagnosis in resource-limited clinical environments.

Background. Missing data undermines data quality and secondary use of clinical databases, including electronic health records (EHRs), registries, and longitudinal cohorts. While numerous imputation methods exist, their comparative performance in databases with bounded clinical variables (e.g., neuropsychological scales, functional assessments, laboratory values with known measurement ranges) remains systematically underexplored. Moreover, most imputation frameworks fail to ensure that recovered values satisfy clinical validity constraints-a critical requirement for downstream analyses.Methods. We conducted a comprehensive benchmarking study of 10 imputation methods (mean/median, KNN, MICE, MissForest, HyperImpute, EM, GAIN, VAEM, VaDER) on a clinical database with 1,357 patient records and 51 bounded variables. We artificially introduced missingness at 10-40% rates under three mechanisms (MCAR, MAR, MNAR) and evaluated both unconstrained and constrained configurations with post-imputation clinical range enforcement. Performance was assessed using accuracy metrics (MAE, RMSE, R²) and clinical feasibility via violation counts (out-of-range predictions).Results. MissForest achieved the lowest error (MAE 2.19±0.02; R²=0.916) across all conditions, substantially outperforming MICE (MAE 2.34, +6.85%), traditional baselines, and deep generative models. Critically, post-imputation constraint enforcement eliminated all clinical violations (MICE: 894→0 violations) with negligible accuracy loss (ΔM AE<0.03), while paradoxically improving performance for naive methods (Mean: −0.17 MAE units). Deep learning methods (VAEM, GAIN) underperformed despite increased complexity.Conclusions. This work establishes a rigorous, evidence-based framework for selecting and deploying imputation methods in clinical databases with bounded variables. MissForest with post-imputation clinical constraint enforcement provides the optimal balance of accuracy and clinical validity, making it immediately applicable to EHRs, registries, and large-scale cohort research. The finding that constraint enforcement is zero-cost and universally beneficial shifts clinical data management practice toward more rigorous secondary-use standards. Code and reproducible benchmarking pipeline are provided.