Research & Innovation
Original Research
AlphaX researchers publish original work in international journals: medical AI, federated and edge AI, neuroscience and space systems.
Published papers by journal ranking. Each paper is counted once.
6
Papers published
Accepted or published
1
D1
Top 10% of journals
4
Q1
Top 25% of journals
1
Q2
Second quartile
6
Public projects
0
Under review
0
In progress
4.35
Average impact factor
What do D1, Q1 and Q2 mean?
Journals in a field are ranked by impact and split into four quartiles. Q1 is the top 25%, Q2 the next 25%. D1 is the top 10%, the strongest part of Q1. Each paper below shows its journal's quartile and impact factor.
LunarFL: Federated Terrain Learning Across Distributed Planetary Surface Assets Under Intermittent, Thermally-Gated Connectivity
As the Artemis and Moon-to-Mars programs advance toward a sustained lunar presence, semi-autonomous surface assets must perceive terrain onboard while sharing a scarce, thermally constrained link budget. Federated learning (FL) fits this setting: assets improve a shared perception model by exchanging updates rather than raw imagery. Existing FL studies, however, assume near-persistent connectivity and evaluate on synthetically partitioned data. This paper reports what is, to the author’s knowledge, the first evaluation of federated terrain segmentation on real planetary rover imagery under a physically motivated model of intermittent, thermally-gated connectivity. A naturally non-independent-and-identically-distributed (non-IID) three-client federation is built from the AI4Mars Curiosity dataset and the synthetic Artificial Lunar Landscape dataset; its heterogeneity is measured, not assumed (mean pairwise Jensen-Shannon divergence 0.42). A SegFormer-B0 model is trained with FedAvg while an aggregation probability pagg and a thermal duty cycle gate every round. Over 113 runs with up to five seeds per condition, accuracy declines monotonically with aggregation scarcity: federation recovers 88.8% of the gap between local-only (0.363 mIoU) and centralized (0.897) training at pagg = 1, 78.6% at 0.3, and 5.9% at 0.05. The deficit scales as paggb with b = −0.650 (95% CI [−0.755, −0.545]), steeper than the optimization-limited 1/√pagg reference and consistent with heterogeneity-driven client drift. Neither a proximal term nor server momentum recovers the scarce regime, and the temporal structure of outages, not the mean rate alone, governs predictability. Uplink volume is reported alongside accuracy at every operating point.
MDEF-AI: a multi-level dynamic evaluation framework for assessing the impact of AI-based clinical decision support on decision quality and patient safety
Background Artificial intelligence-based clinical decision support systems (AI-CDSS) are increasingly deployed in healthcare settings, yet their impact on clinical decision quality and patient safety remains poorly understood. The dominant research paradigm evaluates AI-CDSS through algorithmic performance metrics, accuracy, sensitivity, and specificity, that do not capture the cognitive, behavioral, and organizational dynamics through which AI recommendations are translated into clinical decisions. This performance-centric orientation creates a persistent translation gap between demonstrated technical capability and achieved clinical benefit. Objective To propose and theoretically ground a multi-level, dynamic conceptual evaluation framework (Multi-Level Dynamic Evaluation Framework for AI-CDSS; MDEF-AI) that integrates system design characteristics, human-AI interaction mechanisms, contextual moderators, and temporally dynamic feedback loops to provide a comprehensive analytical lens for assessing the impact of AI-CDSS on clinical decision quality and patient safety. Methods A literature review and critical analysis methodology was employed, involving review of 52 peer-reviewed publications across eight thematic domains, thematic analysis, critical evaluation of existing frameworks, and conceptual synthesis. The resulting framework was systematically compared against the three most closely related existing models to establish its distinct theoretical contribution. Results MDEF-AI is a three-level evaluation structure comprising: (1) contextual moderators across four categories, clinical context, clinician characteristics, organizational factors, and patient characteristics; (2) a core dynamic chain of five interdependent components, system characteristics, human-AI interaction mechanisms (encompassing six mechanisms), clinician cognitive and behavioral responses, clinical decision quality, and patient safety outcomes (encompassing five dimensions); and (3) three temporally distinct feedback loops. AI Omission Trust, the tendency to treat the absence of an AI alert as active clinical reassurance, independent of knowledge of the system’s false negative rate, is proposed as a conceptual interaction failure mode that is hypothesized to operate passively, to resist the design interventions effective against automation bias, and to generate patient safety harm that is largely invisible to existing incident reporting systems. Critical analysis identified six structural gaps in the literature, each corresponding to a specific MDEF-AI design element. Conclusion Patient safety in the age of clinical AI is an emergent property of the entire sociotechnical ecosystem, not a product of algorithmic performance alone. MDEF-AI provides the theoretical architecture for evaluating AI-CDSS in a way that accounts for this complexity, and generates a structured agenda for empirical validation, instrument development, governance reform, and longitudinal evaluation research.
Montage-agnostic federated learning for privacy-preserving Alzheimer’s disease classification from heterogeneous multi-site EEG
Federated learning (FL) lets institutions train shared models without moving raw data, attractive for clinical electroencephalography (EEG), where cohorts are small and privacy-sensitive. Two obstacles persist: clinical sites record EEG with different electrode montages, so input feature spaces differ and standard aggregation fails; and because one recording yields thousands of epochs, splitting epochs rather than subjects leaks subject identity and inflates accuracy. We present MontageFL, a montage-agnostic pipeline that maps any electrode configuration to a fixed 88-dimensional feature vector through region-based aggregation of spectral and covariance features, letting a 19-channel and an 11-channel site train together. Benchmarking centralized, local, FedAvg, and FedProx training for Alzheimer’s disease versus healthy-control classification (Miltiadous et al. dataset; 65 subjects) under strictly subject-level, leakage-free evaluation, subject-level accuracy reached approximately 74% (19-channel) and 89% (11-channel), with no detected difference between federated and local models (FedProx versus local, pp; 95% CI ), consistent with parity but unable to exclude a modest effect given the small cohort. Evaluated with epoch-level splitting, the same pipeline scored 15 to 30 points higher, illustrating leakage-driven inflation. Because montage heterogeneity was simulated by channel subsetting within one acquisition system, validation on independent recording hardware remains future work.
Fine-tuned ResNet34 for efficient brain tumor classification
Brain tumors are among the most fatal diseases, Often leading to a reduction in life expectancy. Early and accurate diagnosis is essential to guide effective treatment and enhance survival rates. Advances in artificial intelligence, particularly deep learning with Convolutional Neural Networks (CNNs), have revolutionized medical imaging analysis by enabling automated and precise diagnostic tools. This study investigates the effectiveness of deep transfer learning for brain tumor classification using a publicly available 7023 image Brain Tumor MRI Dataset (Figshare, SARTAJ, Br35H), and split into training, validation, and test sets, categorizing four classes : glioma, meningioma, pituitary tumor, and no tumor. The proposed method employs a fine-tuned ResNet-34 model, enhanced with custom classification head, data augmentation techniques, and the Ranger optimizer (combining RAdam and Lookahead for stable convergence). The model achieved an accuracy of 99.66% , surpassing current state-of-the-art approaches.
Translational gaps and clinical readiness of artificial intelligence and multimodal imaging in breast cancer diagnostics
Background A persistent translational gap separates the high research-benchmark performance of artificial intelligence (AI) and advanced imaging in breast cancer from demonstrated real-world clinical utility. Most systems reporting accuracy above 95% on curated datasets have not been externally validated across diverse populations and imaging platforms. Purpose We critically examine AI-driven diagnostics and advanced imaging modalities for breast cancer, focusing on the barriers, generalization failure, algorithmic bias, reproducibility, and workflow integration, that determine clinical readiness rather than on benchmark capability alone. Methods We conducted a structured review of peer-reviewed literature (January 2015, March 2025) across PubMed, Scopus, IEEE Xplore, and Web of Science, combining breast cancer terms with AI, deep learning, vision transformers, digital breast tomosynthesis (DBT), contrast-enhanced mammography (CEM), MRI, and liquid biopsy. Study selection followed an explicit, PRISMA-style screening workflow; 127 of 312 identified records met predefined eligibility and quality criteria. Results Reported benchmark performance is consistently optimistic relative to clinical performance: deep learning exceeds 95% accuracy and vision transformers report near-perfect classification on CBIS-DDSM, yet a reproducible 5 to 15% degradation follows deployment owing to distribution shift, scanner and vendor heterogeneity, and population differences. Among validated technologies, DBT raises cancer detection by 27% while lowering false positives; CEM increases sensitivity from 71.5% to 92.7% in dense breasts; and the Mirai risk model (AUC 0.76−0.81) outperforms traditional clinical tools. Liquid biopsy offers high-specificity molecular profiling but inadequate sensitivity (33%) for population screening. Critically, technical maturity is decoupled from validation depth and equity readiness, the dimensions that actually gate adoption. Conclusion Realizing the potential of AI-augmented, risk-stratified screening requires multicenter prospective and multi-vendor external validation, fairness-aware design with subgroup reporting, transparent reproducibility, and adaptive regulation of continuously learning systems. We propose an operationally defined translational readiness framework that scores each technology along six dimensions, locating it on the continuum from research capability to equitable clinical deployment.
A conceptual multi-agent architecture for mental health triage in post-conflict Arabic-speaking populations: a theoretical proposition and staged validation argument
Syria’s protracted conflict has produced a mental health crisis of extraordinary scale, with post-traumatic stress, depression, and anxiety estimated at several times global baselines, set against fewer than 0.37 psychiatrists per 100,000 people. Existing AI mental health tools have been developed and evaluated primarily for English-speaking, non-humanitarian populations, and their transfer to this setting is constrained by three simultaneous structural deficiencies, extreme clinical scarcity, Arabic natural-language-processing underperformance for dialect, and cultural misalignment with Syrian idioms of distress, which we term the Triple Gap. This article is a conceptual contribution in the Hypothesis and Theory genre, and its central claim is theoretical rather than technical: that AI-assisted mental-health triage at a safety floor adequate for crisis relevant care in this setting is conditional on the joint satisfaction of three constraints, linguistic adequacy for the local dialect, cultural validity for local idioms of distress and help-seeking, and bounded clinical responsibility through human oversight. These constraints interact, so that a system satisfying fewer than all three is expected to fail in clinically consequential rather than random ways. As one possible design response to this proposition, neither the only one nor a validated one, we describe a conceptual multi-agent architecture aligned with the WHO mhGAP task-shifting model: a four-stage pipeline (screening, risk stratification, routing, follow-up) constrained by a cross-cutting cultural-adaptation layer, augmented by candidate verification mechanisms with explicit abstention, and governed by human oversight in which clinical responsibility rests with a licensed clinician. The proposal is a hybrid clinical decision support hypothesis, not an autonomous system. Because no Syrian Arabic clinical corpus yet exists, the conversational components of the design cannot presently be evaluated; we therefore set out a staged sequencing argument for future work in which the construction of a Syrian Arabic Mental Health Evaluation Corpus (SAMHEC) is the first and rate-limiting condition. We present no prototype, no corpus, and no clinical, cultural, or safety evaluation, and we make no claim of clinical validity, safety, or readiness for deployment. The contribution is the integration of multi-agent triage with task-shifting and cultural adaptation into a single conditional argument whose adequacy can be established only through the staged empirical work we describe.