Research

On Device Small Language Models for Symptom Diary P.R.O. Measures

Academic Context

Flarelines is developed as a research artifact in the master's studies in Speculative Software Design at the University of Applied Sciences Potsdam. The research explores AI-assisted health journaling for chronic illness patients with a focus on privacy.

The Problem

When people with chronic health conditions like systemic lupus erythematosus (SLE) attend routine medical visits, the practitioner receives only a snapshot of the patient's current state. Symptom flares that occurred in the past are difficult to recall accurately.

Free-text symptom diaries help, but they're often too verbose for clinicians to interpret within limited appointment times. A more structured representation is needed.

The Approach

We use small language models (SLMs) that run entirely on-device to transform free-text journal entries into structured scores:

This shifts clinical NLP away from remote API services, which introduce privacy risks. On-device processing keeps health data private while producing clinically useful output.

The Model

We fine-tuned Qwen2.5-3B on lupus diary data using a hybrid approach: autobiographical journal entries from a lupus patient (one of the researchers) combined with synthetic data generated by GPT-5.1.

The model runs locally on iPhone (~1.8 GB). No internet connection required for scoring.

Data Contribution

Users can optionally contribute anonymized journal entries (text + scores only) to improve the model. No identifiers are collected. This is disabled by default and can be enabled in Settings.

Contact

Research inquiries: flarelines@inpyjamas.dev