AutoMedic: Automated Medical Systems

• AutoMedic is a suite of automated medical systems that integrate AI, embedded devices, and ML to monitor vitals, manage medication, and automate diagnostics.
• The framework spans wearable patient monitoring, pill dispensing reminders, conversational AI evaluation, and automotive fault analysis using specialized hardware and algorithms.
• Practical deployments demonstrate real-time alerting, secure data transmission, and high diagnostic accuracy, while highlighting limitations and opportunities for system enhancements.

AutoMedic refers to a suite of automated, embedded, or AI-driven systems for medical and diagnostic applications, spanning clinical wearable devices, smart assistance for medication, medical conversational AI evaluation, and even extensions to automotive diagnostic automation. The term encompasses diverse implementations, ranging from microcontroller-based vital monitoring and automatic pill reminders to advanced machine learning frameworks for clinical conversation evaluation and automotive fault diagnosis, each distinguished by its specialized hardware, algorithmic, and deployment strategies.

1. Wearable Patient Monitoring and Automated Alerting Systems

The Medisûr AutoMedic system exemplifies microcontroller-based wearable healthcare automation (Chatterjee et al., 2019). Its architecture integrates an ATmega328-driven wristband measuring pulse (via a PPG-based M212 sensor), body temperature (LM35), and emergency button state, transmitting encrypted (AES-256) readings via 434 MHz RF to a receiver node. The receiver, incorporating Arduino Uno, GPS (NEO-6MV2), and GSM (SIM300), decrypts the payload, augments it with live geolocation, and uploads the encrypted, time-stamped context to a Microsoft Azure-hosted backend over GPRS.

Upon detecting abnormalities (temperature >99°F or HR<40 bpm) or button presses, the system dispatches real-time SMS alerts containing device ID, vitals, and location. Cloud infrastructure built around Azure VMs and SQL Server enables live data tracking and historical display via a cross-platform mobile app. Prototype performance demonstrates ~15 s total alert latency, with sensor accuracy of ±0.5°C for temperature and ±1 bpm for heart rate; the design supports 1 min reporting intervals and expects 8–12 h operational lifetime on a 200 mAh Li-Po (assuming future sleep-mode optimizations).

Key limitations include lack of direct SpO₂/blood pressure measurement, unidirectional and unprotected physical-layer RF communication, reliance on a global AES key, and fully specified error handling only as recommended future work. The proposed architecture outlines future extensibility towards higher-fidelity sensing (e.g., ECG), EHR interoperability, and secure wireless protocols.

2. Automated Medication Dispensing and Reminder Devices

AutoMedic, as realized in the pill reminder and dispensing context, implements time-based medication adherence assistance using commodity electronics (Jabeena et al., 2017). The system is architected around an Arduino Uno, DS1307 RTC, SIM800L GSM for SMS notifications, and IR sensors for compartment status. The software maintains a prescribed dose schedule and activates alarms (LED + buzzer) for the correct compartment at programmed times. If the patient opens the appropriate compartment within a designated interval, the event is logged; otherwise, the system snoozes the alarm or escalates via SMS to the patient and caregiver.

The system design omits motorized dispensing—retrieval is manual but logged via IR beam-break, ensuring simplicity and low component cost. Notification redundancy is achieved by combining audible, visual, and remote (SMS) modalities. Major limitations are the absence of motorization (no pill wheel), restriction to three dosage slots, and lack of formal evaluation metrics (e.g., timekeeping drift, sensor error rates). Suggested future work includes adopting actuated dispensers, IoT-based state upload, and adherence analytics.

3. Automated Evaluation of Medical Conversational Agents

AutoMedic, as an AI evaluation framework, enables multi-faceted assessment of clinical LLM conversational agents in dynamic, simulated multi-turn dialogue (Oh et al., 11 Dec 2025). The architecture orchestrates four LLM-based agents: a profile generator (converting static QA items into virtual patient profiles), a patient, a clinical staff entity, and the doctor agent under evaluation. Each static QA item is transformed to a structured patient profile $P = (D, B, O)$, where $D$ is demographics, $B$ basic symptoms/history, $O$ optional labs/vitals—inferred as needed from context priors $$p(X_j \mid \text{demographics}, \text{chief complaint})$$ with outcome-preserving constraints.

Dialogue is rolled out turnwise, with the doctor agent querying patient or staff, accumulating knowledge until termination and final multiple-choice answer. The CARE metric scores clinical conversational performance across four axes:

• Accuracy ($S_{\mathrm{ACC}}$): Fractional preservation of static QA performance in the multi-turn setting.
• Conversational Efficiency and Strategy ($S_{\mathrm{CES}}$): Inverse mean words per turn if correct; else zero.
• Empathy ($S_{\mathrm{EMP}}$): Patient-agent blinded rating, normalized to [0,1].
• Robustness ($S_{\mathrm{ROB}}$): Proportion of simulations without failure (e.g., role confusion, early exit).

Empirical evaluation on six benchmark QA datasets with major LLMs, both general and biomedical, revealed nontrivial accuracy drop in the conversational setting (Pearson $r = 0.7547$), distinct agent profile strengths across CARE sub-domains, and strong alignment of CARE with human expert ratings. Practical insights include the need for balanced fine-tuning that preserves both domain knowledge and conversational skill, the primacy of certain datasets (MedBullets, MedQA) for simulation fidelity, and the urgency of developing multimodal (vision + text) extensions.

4. Medical Assistant Systems on Edge Devices

Modern AutoMedic systems adopt on-device, multi-agent architectures to provide privacy-preserving, latency-optimized healthcare services, including appointment booking, vitals monitoring, reminders, and daily reporting, independent of cloud infrastructure (Gawade et al., 7 Mar 2025). A typical instantiation leverages a quantized large action model (Qwen2.5-Coder-7B-Instruct, split into Planner and Caller agents via LoRA) to sequentially plan and execute user tasks. The Health Manager orchestrates background logging and threshold-based vitals alerting, while the Scheduler extracts and computes adherence timings from parsed prescriptions.

Short- and long-term memory (STM/LTM) are maintained in encrypted SQLite, with context/record retrievals vectorized using spaCy. Task allocation aims to minimize cumulative agent latency under device memory constraints, and all inference and data management occur locally. The system achieves BLEU and ROUGE-L scores of 0.84/0.855 (planning), 0.99/0.965 (calling), and supports sub-500 ms end-to-end response on commodity smartphone CPUs. Key design tradeoffs address model modularity, memory footprint (<5 GB for full deployment), and user data governance.

5. Laboratory Automation: White Blood Cell Classification

AutoMedic also comprises compact diagnostic hardware for point-of-care analysis, as illustrated by the automatic WBC measuring aid (Ghosh et al., 2013). This implementation integrates an imaging pipeline—microscope + CCD, microcontroller-driven XY stage, and onboard image processing—that performs Laplacian edge sharpening, HSI color space segmentation (using iterative thresholding on hue), morphological noise removal, overlap separation, and geometric/color feature extraction.

Classification into five major WBC types uses aspect ratios, center-of-mass offset, and cytoplasm color via fuzzy membership functions. On 150 validation samples, the system achieved 97.33% overall accuracy, with class-wise F1 scores >95%. Embedding and reporting modules support telemedicine via Wi-Fi/Ethernet, enabling remote review. This design directly addresses settings with no access to reference laboratories or skilled personnel, prioritizing robustness and field usability.

6. Diagnostic Automation in Automotive and Multilingual Fault Texts

In automotive contexts, the "AutoMedic" designation extends to both sound-based diagnostics and multilingual textual fault classification. The OtoMechanic system demonstrates audio-based fault identification, using log-Mel spectrograms and VGGish CNN embeddings, yielding oracle top-1 and top-3 accuracies of 58.7% and 82.9% respectively on 12-way vehicle fault classes (Morrison et al., 2019). The workflow combines category metadata with query-by-example audio retrieval and supports real-time interactivity.

Separately, automotive fault diagnosis from textual claims leverages multilingual Transformer models (XLM-R), attaining >80% top-1 accuracy on the most frequent classes, with robust adaptation across 38 languages and 1,357 root-cause categories (Pavlopoulos et al., 2022). The pipeline, combining language detection, tokenization, and large-scale fine-tuning, demonstrates high performance in dense classes and supports integration with sensor and rule-based diagnostic signals via log-linear fusion.

7. Limitations and Prospective Directions

Across modalities, prominent AutoMedic system limitations include limited evaluation on large-scale or longitudinal data (especially for adherence/pill reminder hardware), incomplete closed-loop security (notably global key usage and non-TLS RF transmission), and the absence of hardware miniaturization or compositional integration with EHR and telemedicine platforms. In LLM evaluation, current frameworks are restricted to text; extending agentic simulations to include imaging or clinical waveform data remains an open challenge. Future work recommends modular, upgradable architectures, attention to context-adaptive ML-driven reminders, formal user studies for efficacy and equity, and continued development of clinically calibrated composite metrics for conversational performance and safety.