💻 computer science

AlamX: A Privacy-Preserving Local AI Framework for Symptom-Based Disease Prediction and Conversational Clinical Triage

AlamX is a privacy-preserving, fully local clinical intelligence platform that combines a Random Forest-based symptom and vital triage system with a locally executed biomedical LLM to provide accurate, uncertainty-aware disease predictions and conversational health guidance without transmitting sensitive user data to remote servers.

Original authors: Hamid Alam, Shubham Kumar, Rahat Kr Pradhan, Saksham Pradhan, Rohit Subba, S D Mohana

Published 2026-09-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Hamid Alam, Shubham Kumar, Rahat Kr Pradhan, Saksham Pradhan, Rohit Subba, S D Mohana

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern world, checking a symptom often means turning to a screen. A person feels a rash, a headache, or a strange pain, and they type it into an app or a website. These digital tools promise to offer a quick answer, a possible name for what is wrong. But to get that answer, the user usually has to send their personal health details across the internet to a remote server. This creates a difficult choice: to get help, one must surrender privacy. The data that makes a diagnosis accurate is also the data that makes a person most vulnerable if it is intercepted or leaked. Furthermore, most of these tools act like a single-lane road. They take a list of symptoms and spit out one disease name, acting as if the answer is certain, even when the signs are vague. They often ignore other clues a person might have, like their heart rate or how well they slept, treating the complaint in isolation from the rest of the body.

A team of researchers at the Sikkim Manipal Institute of Technology in India has built a different kind of tool to solve this problem. They call it AlamX. Instead of sending health data to a distant server, AlamX lives entirely on the user's own computer or device. It is a complete system that runs locally, meaning no information leaves the machine. The researchers designed it to take a list of symptoms, compare them against a vast library of medical patterns, and offer a diagnosis that includes a measure of how sure the system is. It also listens to the user's questions in a natural conversation, using a specialized artificial intelligence that understands medical terms but never needs to connect to the internet to think. The goal was not just to make a smart program, but to make a private one that respects the user's data while still providing useful medical guidance.

The system works by asking the user to select from a standardized list of 132 possible symptoms, such as a skin rash, a fever, or swollen joints. Once the user makes their choices, the software converts these selections into a simple code and runs them through a decision-making engine. This engine, trained on thousands of examples, looks for patterns that match 41 different diseases. Unlike many other tools that simply guess one answer, AlamX provides a primary prediction along with a percentage that shows how confident it is in that guess. It also lists other possible conditions that share similar symptoms, giving the user a clearer picture of the possibilities. If the user is unsure, they can ask the system questions in plain language. The system answers by looking at the specific symptoms and health data the user just entered, all without ever transmitting that information to a cloud server.

To make this work on a personal device, the researchers had to overcome several hurdles. They needed a way to handle the fact that some diseases are much rarer than others in the training data, which can trick a computer into ignoring the rare ones. They used a technique to balance the data so the system learns to recognize both common and uncommon conditions. They also built a "health score" that combines the user's daily steps, heart rate, and sleep patterns with their symptoms, creating a single number that tracks their overall well-being over time. This score updates continuously, offering a view of the user's health that changes as they do, rather than just a snapshot of a single moment.

The results of building this system show that the diagnostic engine produces a confusion matrix with a dominant diagonal across all 41 classes, indicating that predictions concentrate on the true class, though specific accuracy values have not yet been reported. In a specific test case, a user selected two symptoms: a skin rash and itching. The system responded with a primary prediction of a fungal infection, stating it was 71 percent confident. It did not stop there; it also suggested a drug reaction as a secondary possibility with 27 percent confidence. This behavior is crucial because it acknowledges that two symptoms can point to more than one thing, and it tells the user exactly how much the system knows and how much it is guessing.

The researchers also integrated a large language model, a type of artificial intelligence trained specifically on medical texts, to handle the conversation. This model runs locally on the user's machine, meaning it can answer questions about the diagnosis or explain what a symptom might mean without sending a single word of the user's health history to an external provider. The entire system, from the database that stores the user's history to the intelligence that answers questions, resides on the device itself. This design ensures that even if the internet goes down, the tool still works, and no one else ever sees the data.

While the system shows great promise, the authors are careful to note its limits. The tool is designed to help people decide when to seek care, not to replace a doctor. It works with a fixed list of 41 diseases and relies on the user to input their symptoms and vital signs manually. The researchers did not test the system on real-world patients in a hospital setting, so its performance in a live clinical environment is still unknown. They also noted that the system currently treats symptoms as simple on-or-off switches, rather than accounting for how severe or long-lasting a symptom might be. Despite these limitations, the work demonstrates a significant shift in how health technology can be built. It proves that a system can be accurate, aware of its own uncertainty, and capable of conversation, all while keeping the user's private data completely safe on their own device.

The broader lesson from this project is that privacy and utility do not have to be enemies. For years, the assumption has been that to get the best medical advice from an app, one must give up control of personal data to a central server. AlamX challenges that idea by showing that powerful analysis can happen right where the data is created. The system does not need to be a magic box that knows everything; it is a practical tool that respects the user's boundaries. By keeping the processing local, it removes the risk of data breaches and allows people in remote areas or with limited internet access to get immediate, informed guidance. The researchers have built a framework that could be expanded in the future to include more diseases and more detailed health tracking, but the core achievement is already clear: a private, local, and intelligent assistant for health is not just a theoretical idea, but a working reality.

Technical Summary: AlamX – A Privacy-Preserving Local AI Framework

1. Problem Statement

Current symptom-based disease prediction systems face four critical architectural limitations that hinder their utility in privacy-sensitive and resource-constrained settings:

Privacy Exposure: Most systems require transmitting sensitive personal health information (symptoms, vitals, lab values) to centralized cloud infrastructure, creating inherent disclosure risks.
Symptom-Only Inference Ceiling: Many consumer tools ignore physiological vitals, despite evidence that symptom-only inference has a significantly lower performance ceiling compared to multimodal approaches.
Uncalibrated Outputs: Existing systems typically return a single, unqualified disease label, failing to express diagnostic uncertainty or provide differential diagnoses when symptom presentations overlap.
Prototype-Deployment Gap: Many implementations remain as academic notebooks or simple GUIs (e.g., Tkinter, Gradio) lacking authentication, persistent user state, encrypted credentials, or documented API contracts.

The core challenge is to build a system that is simultaneously accurate, uncertainty-aware, capable of fusing symptoms with vitals, and entirely local to the user's host.

2. Methodology and System Architecture

AlamX is a full-stack, two-tier clinical intelligence platform where all analytical components execute locally on the user's device. The system is decoupled into four cooperating tiers:

Architecture

Presentation Layer: A React single-page application (served by Vite) that renders within a mobile-emulating container. It manages session state via localStorage and handles user interaction for symptom selection and chat.
Application Layer: A FastAPI service (served by Uvicorn) acting as the central logic hub. It manages five routers: authentication, vitals, symptoms, prediction, and chat. It enforces security via bcrypt password hashing and signed JSON Web Tokens (JWT).
Data Layer: A file-backed SQLite database provides relational persistence for users, vitals, symptom logs, predictions, and chat history. It ensures transactional integrity without exposing a network-accessible database daemon.
Inference & Language Model Layers:
- Diagnostic Engine: A pre-trained Random Forest classifier (Scikit-Learn) loaded at startup for low-latency inference.
- Conversational Engine: The BioMistral biomedical large language model (LLM) hosted locally via the Ollama runtime.

Data Processing and Modeling

Input Standardization: Users select symptoms from a standardized vocabulary of 132 clinical parameters. These are encoded into a 132-dimensional binary feature vector ( $x \in \{0, 1\}^{132}$ ).
Class Imbalance Correction: The training pipeline applies SMOTE (Synthetic Minority Over-sampling Technique) exclusively to the training partition to address class imbalance without leaking synthetic data into the evaluation set.
Model Training: A Random Forest is trained as the primary deployment model due to its suitability for sparse, high-dimensional binary data and its intrinsic ability to estimate probabilities via vote distribution. An XGBoost model is trained in parallel as a comparative baseline.
Health Score Engine: A longitudinal pathway fuses daily vitals (footsteps, heart rate, sleep) with active symptom logs to generate a composite wellness index. This engine uses a non-decreasing mathematical property to ensure that logging healthy activities never penalizes the user's score.
Conversational Grounding: The chat router assembles structured health context (vitals, recent predictions, logged reports) and injects it into prompts sent to the local BioMistral model, enabling context-aware reasoning without external transmission.

Output Generation

The system returns a multi-faceted inference artifact derived from a single probability distribution:

Primary Prediction: The disease class with the highest probability.
Confidence Score: A percentage derived from the ensemble vote distribution.
Differential Signals: A ranked list of alternative diagnoses (top- $k$ ) to communicate uncertainty.
Latency & Disclaimer: Measured inference time and an explicit clinical safety disclaimer.

3. Key Contributions

The paper identifies five primary contributions, emphasizing architectural innovation over marginal accuracy gains:

(C1) Fully Local Architecture: Discriminative classification, relational persistence, and generative language reasoning all execute on the user's host, eliminating the need to transmit protected health information (PHI).
(C2) Unified Data Substrate: A single symptom log drives both episodic disease prediction and continuous Health Score calculation, fusing signals without duplicating user effort.
(C3) Uncertainty-Aware Contract: The system outputs confidence, differentials, and latency from a single probability distribution, ensuring that expressing uncertainty imposes no additional inference cost.
(C4) Local Domain-Adapted LLM: Integration of BioMistral under a local runtime with structured health-context injection, replacing the retrieval-based or cloud-dependent chat layers found in prior work.
(C5) Comparative Synthesis: A survey of 16 studies isolates six architectural research gaps (e.g., privacy, symptom-vital fusion, deployment engineering) and maps each to a specific design decision in AlamX.

4. Results and Evaluation

Classification Behavior: Evaluation of the Random Forest over the 41-class label space yields a confusion matrix with a dominant diagonal and negligible off-diagonal mass, indicating strong classification behavior without systematic confusion between adjacent conditions.
Feature Importance: Gain-based analysis reveals that 130 of the 132 available parameters are utilized by the model, suggesting the standardized vocabulary is nearly fully informative. Specific signs like "swollen extremities" and "swollen blood vessels" carry the highest discriminative weight.
Functional Demonstration: A worked inference example (symptoms: skin rash, itching) returned "Fungal infection" as the primary prediction with 71% confidence and "Drug Reaction" as a differential signal at 27%. This demonstrates the system's ability to communicate genuine diagnostic uncertainty rather than suppressing it.
Security: The system enforces data locality; the database, model artifacts, and LLM weights reside entirely on the deployment host. No outbound network connection is required for core functionality.

Note: The paper explicitly states that quantitative metrics (accuracy, F1, latency means/variances) and clinical validation against real-world datasets were not specified in the project documentation and remain future work.

5. Significance and Claims

The paper argues that the primary barrier to effective symptom-based triage is architectural, not algorithmic. While existing literature shows that classification accuracy on standardized corpora is near saturation, the utility of these systems is compromised by privacy risks and opaque outputs.

AlamX demonstrates that:

Privacy and Utility are not competing objectives: A system can be highly accurate and context-rich while remaining entirely local, inverting the traditional privacy-utility trade-off.
Local execution enables richer context: By keeping data local, the system can fuse detailed vitals and history into the LLM's reasoning without the disclosure risk associated with cloud transmission.
Resilience: The platform remains functional without internet connectivity, making it viable for resource-constrained or intermittently connected settings.

The authors conclude that the significance of this work lies in proving that a privacy-preserving, local execution model can deliver clinically useful analytical capabilities, including uncertainty-aware outputs and conversational grounding, without relying on centralized infrastructure.