Mobile Disease Diagnostics

• Mobile disease diagnostics are systems that leverage mobile hardware, AI, and sensor fusion to enable on-site detection and monitoring in resource-constrained environments.
• They deploy diverse techniques including imaging analysis, biosensing, and molecular testing to provide rapid, accurate, and accessible diagnostic services.
• Advanced methodologies like lightweight CNNs, quantization, and federated learning enhance performance, scalability, and data privacy in dynamic field conditions.

Mobile disease diagnostics refers to the deployment of computational, sensor, and analytic capabilities on mobile devices—primarily smartphones and tablets—for the detection, screening, and monitoring of human, animal, and plant diseases. Leveraging advancements in embedded artificial intelligence, imaging hardware, bio-sensing modalities, and mobile networking, this paradigm is defined by on-site, often offline, disease inference workflows operable in resource-constrained settings. Mobile disease diagnostics aims to democratize access to expert-level diagnostic capabilities, with applications ranging from dermatological screening to infectious disease surveillance, chronic disease risk scoring, molecular detection, point-of-care laboratory tests, and agricultural pathology.

1. System Architectures and Core Algorithmic Components

Mobile disease diagnostics solutions occupy a diverse spectrum of system architectures, determined by sensing modality (imaging, chemical, acoustic, physiological), on-device versus cloud inference, user interface requirements, and operational environment.

• Imaging-based clinical screening: Pipeline designs, as exemplified by AICOM-MP (Yang et al., 2023), integrate (i) image acquisition via smartphone camera, (ii) pixel normalization, (iii) on-device neural network inference (MobileNetV2 or EfficientNet-Lite0), (iv) softmax-based class probabilities, and (v) post-processing with confidence calibration and recommendation logic. Image-based diagnostics are leveraged for classification and triage in dermatology, ophthalmology, chest radiographs, and plant pathology.
• Laboratory test interpretation: Mobile platforms support digitization and AI-based reading of rapid diagnostic test (RDT) kits, using object detection (YOLOv8) to localize membrane regions and convolutional neural networks for result interpretation, with significant accuracy gains and accessibility enhancements for visually impaired users (Dastagir et al., 2024).
• Biosensing: Fully integrated electrochemical potentiostat interfaces (Fu et al., 2015), magnetic particle spectrometers (MPS) (Wu et al., 2020), and interfacial-potential transduction readers (Jang et al., 24 Mar 2026) transform smartphones into universal biochemical analysis tools by providing Bluetooth or USB connectivity, analog front-ends, and custom mobile applications for workflow automation, measurement, and cloud synchronization.
• Molecular diagnostics: Platforms like "DNA-to-go" (Priye et al., 2016) combine thermal cycling (via convection PCR), smartphone-based fluorescence imaging, and custom apps for real-time data analysis. Isothermal assay platforms (e.g., RPA-CRISPR-Cas12a with microneedle sampling (Zhao et al., 10 Jun 2025)) enable nucleic acid detection using fluorescent readouts captured by mobile phone cameras.
• Federated and privacy-preserving analytics: For population-scale mobile health diagnostics, cross-device federated learning frameworks (e.g., FedLoss (Xia et al., 2023)) aggregate distributed, non-IID data from edge devices without centralizing sensitive user data.
• Multimodal chronic disease risk prediction: Mobile- and web-integrated large language multimodal models (LLMMs) (Liao et al., 2024) fuse text from clinical notes and structured lab values, enabling risk stratification for hypertension, diabetes, and cardiovascular disease directly on physician-facing mobile apps.

2. Model Architectures, Compression, and Performance Benchmarks

Mobile diagnostics demand models that balance diagnostic performance with stringent compute, memory, and energy budgets.

• Lightweight CNNs and transformers are predominant for image classification tasks. Example: MobileNetV2 (2.3M parameters, ≈300M FLOPs), EfficientNet-Lite0 (4.7M parameters) (Yang et al., 2023), and MobilePlantViT (0.69M parameters, 0.3–0.6 GFLOPs per inference) (Tonmoy et al., 20 Mar 2025). For multi-class classification in plant pathology, accuracy up to 99.57% is achievable (PlantVillage dataset) (Tonmoy et al., 20 Mar 2025).
• Compression techniques: Structured filter pruning and post-training 8-bit quantization routinely shrink model size by 50–75% with <1.5% absolute loss in classification accuracy (Yang et al., 2023, Foysal et al., 2024). In TensorFlow Lite deployments, quantized models can achieve <200 ms inference latency, <10 MB RAM, and per-sample energy usage <0.05 J.
• Ensemble and hierarchical models enable multimodal assessment, as in attentive deep learning aggregators for smartphone-based Parkinson’s diagnosis (AUC=0.85) (Schwab et al., 2018) and LLMMs for chronic disease (Liao et al., 2024).
• Non-vision models: Federated learning aggregates client-unique models using client-wise predictive loss (FedLoss), achieving centralized-level AUC-ROC (0.79) and markedly improved sensitivity versus FedAvg/FedProx (e.g., SE=0.50, SP=0.90, SE@80%SP=0.62 for COVID-19 detection) (Xia et al., 2023).

Model/Platform	Params/Model Size	Accuracy/AUC	Inference Latency	Notable Metric(s)	arXiv ID
MobileNetV2 (AICOM)	2.3M; ~600 kB	93.2% (MPX)	80–120 ms	Sens: 91.5%; AUC: 0.97	(Yang et al., 2023)
MobilePlantViT	0.69M; ~3–8 MB	80–99.57%	40–60 ms (INT8)	Macro F1: 0.9943	(Tonmoy et al., 20 Mar 2025)
DeepLabv3 (TST App)	–	Dice: 0.88–0.90	–	MAE: 0.12 mm	(Gele et al., 22 Jun 2025)
FedLoss FL (COVID-19)	–	AUC: 0.79	–	Sens: 0.50; SP: 0.90	(Xia et al., 2023)
MAIScope (Malaria)	2.2M/7M TFLite	89.9%	25–500 ms	AP: 61.5% (detection)	(Sangameswaran, 2022)

3. Domain-Specific Use Cases and Clinical/Epidemiological Validation

Mobile disease diagnostics span broad application domains, each with unique requirements:

• Idiosyncratic infectious disease identification: AICOM's monkeypox screening pipeline (93.2% accuracy, sensitivity 91.5%, AUC 0.97) demonstrates the feasibility and accuracy of AI-driven image-based clinical triage on mobile hardware (Yang et al., 2023). The platform is designed to be modular and disease-agnostic, supporting extension to malaria (microscopy images), tuberculosis (CXR), and diabetic retinopathy (retinal fundus) via retraining and pipeline modification.
• Tuberculosis Mantoux testing: The mobile TST app employing DeepLabv3-ResNet50 segmentation achieves MAE down to 0.11–0.23 mm and sensitivity/specificity of 96%/94% for positive TST (≥10 mm), with no significant measurement bias versus clinical gold standard (Gele et al., 22 Jun 2025).
• Population-scale, privacy-preserving COVID-19 diagnostics: Cross-device federated learning on multimodal mobile data (acoustics+symptoms) achieves AUC-ROC equivalent to centralized models (0.79), and FedLoss specifically mitigates both local and global class imbalance (Xia et al., 2023).
• Automated malaria microscopy: MAIScope achieves 89.9% RBC-level classification accuracy and 61.5% detection AP entirely offline on a portable platform; AP exceeds comparable YOLO baselines (Sangameswaran, 2022).
• Chronic disease risk scoring: LLMM-based mobile risk prediction platforms ingest EHR text and labs, informing hypertension (F1≈0.33), heart disease (F1≈0.83), and diabetes (F1≈0.70) in real time on mobile and web client UIs (Liao et al., 2024).
• Plant disease management: Pipelines leveraging CNN/ViT architectures with mobile deployment capabilities give leaf disease identification with 98–99% accuracy; YOLOv8-based detection combined with cloud-based transformer classification enables both multi-disease patch detection and server-based treatment lookup (Khanal et al., 2024, Tonmoy et al., 20 Mar 2025).

4. Hardware, Sensing Modalities, and Field Integration

Mobile disease diagnostics platforms integrate or interface with diverse hardware and biosensing modalities optimized for decentralized, resource-constrained environments:

• Imaging: On-device camera sensors (smartphone, Raspberry Pi) combined with depth or AR capabilities (Google ARCore) facilitate clinical and biological measurement; bead-microscope optics yield high-magnification, low-cost imaging for malaria diagnosis (Sangameswaran, 2022).
• Electrochemical analysis: Portable potentiostat modules (potentiostat, microcontroller, Bluetooth stack) enable mobile devices to read cyclic voltammetry for direct biomarker quantification; demonstration with standard redox species and protocol transferability to cardiovascular, infectious, and oncology markers (Fu et al., 2015, Jang et al., 24 Mar 2026).
• Magnetic biosensing: MagiCoil achieves wash-free, one-step, quantitative assays using harmonic analysis of MNP dynamic response, enabling LOD as low as 64 nM for protein biomarkers (Wu et al., 2020).
• Molecular diagnostics: Smartphone-enabled PCR/LAMP platforms integrate compact thermal management (convection heating), fluorescence detection, and intuitive app workflows to deliver 10–20 min pathogen detection with LOD down to several genome copies per reaction (Priye et al., 2016, Zhao et al., 10 Jun 2025).
• AR- and sensor-guided measurement: Mantoux TST application overlays AR circle guides for sticker-based scaling and employs monocular or hybrid depth sensing for geometric calibration (Gele et al., 22 Jun 2025).
• Accessibility and inclusivity: Apps with audio guidance, haptic feedback, and robust detection modules (YOLOv8) enable self-administered testing, including for the visually impaired (Dastagir et al., 2024).

5. Deployment, Security, Scalability, and Limitations

Mobile diagnostics platforms implement operational strategies for scalable, secure, and robust deployment:

• Containerization and orchestration: Systems leverage Docker, Kubernetes, serverless Flask endpoints, and batch queuing (Celery, RabbitMQ) for scalable, concurrent inference and user management (Liao et al., 2024, Khanal et al., 2024).
• Data privacy and compliance: EHRs and user data are de-identified, encrypted in transit and at rest (TLS 1.2+, AES-256), and guarded via OAuth2/JWT RBAC. Compliance with HIPAA-style requirements is engineered via role segregation and audit logging (Liao et al., 2024).
• Offline operation and network independence: Core inference runtimes (TensorFlow Lite, ONNX Runtime Mobile, PyTorch Mobile) provide energy- and memory-efficient local processing, essential in low-connectivity settings; model artifacts are updated periodically when connectivity allows (Yang et al., 2023, Sangameswaran, 2022).
• Model maintenance and monitoring: Retraining pipelines and CI/CD approaches ensure continuous model improvement. Explainability modules (e.g., SHAP, Grad-CAM overlays) provide interpretability for clinician trust and regulatory justifiability (Liao et al., 2024, Dastagir et al., 2024).
• Known limitations: Operational robustness can be affected by lighting variability, skin tone or background bias, and device hardware diversity. Data sets are frequently skewed toward lighter skin or standard imaging, necessitating explicit efforts to increase representativity in training data (Yang et al., 2023, Gele et al., 22 Jun 2025). On very low-end CPUs (<200 MHz), inference latency can exceed practical clinical thresholds; research on further model compression, binarized networks, and edge learning is ongoing (Yang et al., 2023).

6. Generalization, Translation, and Future Directions

Mobile disease diagnostics platforms are intentionally modular and extensible:

• Disease-agnostic core pipelines: Most architectures, after retraining the classifier head or adjusting pre-processing, generalize to new disease domains (e.g., skin lesion → malaria slide → CXR). Modular pre-processing enables adaptation to varied input modalities, such as microscopy versus radiography (Yang et al., 2023).
• Label and data scarcity: Model translation is challenged by insufficient labeled data for rare disorders or emerging diseases; active research targets few-shot learning and self-supervised pretraining for domain transferability (Yang et al., 2023).
• Federated and personalized learning: Cross-device FL with adaptive weighting (FedLoss) enables privacy-preserving continual learning from heterogeneous user data without central aggregation, forming the basis for population-level mobile diagnostic model improvement under real-world label imbalance (Xia et al., 2023).
• Multi-modal and explainable AI: Early integration of multiple sensing modalities (image, symptom, text, lab data) and automated explanation overlays (SHAP for clinical text, attention heatmaps for time-series or image regions) offers clinicians decision support while satisfying requirements for interpretability (Liao et al., 2024, Dastagir et al., 2024).
• Integration with telemedicine: Seamless sharing of results, raw data, and system state with centralized care or epidemiological registries can accelerate public health response, support remote consultation, and enable large-scale surveillance (Liao et al., 2024).
• Regulatory and field validation: Future work cited in multiple studies includes broadening clinical trials, field validation across a range of device types, deployment to endemic regions, and cost-effectiveness analysis (Gele et al., 22 Jun 2025, Yang et al., 2023).

Mobile disease diagnostics constitutes a rapidly evolving interdisciplinary field. The frameworks and workflows described above, supported by robust empirical performance, energy efficiency, and generalizability, illustrate the deployment of advanced AI and sensor fusion into accessible, field-appropriate diagnostic applications across the biomedical, epidemiological, and agricultural domains (Yang et al., 2023, Liao et al., 2024, Dastagir et al., 2024).