| Title |
Wearable data aggregation framework for fine-tuning of LLMs to forecasting future-aware insights |
| Authors |
Petkevičius, Linas ; Gudžius, Povilas ; Filatovas, Ernestas |
| Full Text |
|
| Is Part of |
Proceedings of the 30th international conference on Information Society and University Studies, Kaunas, Lithuania, May 15, 2025.. Aachen : CEUR-WS. 2026, p. 325-334 |
| Keywords [eng] |
LLMs ; AI-based recommendations ; wearable data ; text-based aggregation ; LoRA fine-tuning ; digital health |
| Abstract [eng] |
Wearable sensing and digital health analytics generate large volumes of heterogeneous physiological time series, yet converting these data into accurate and computationally efficient predictive insights remains challenging. While large language models (LLMs) are increasingly applied to health-related dialogue and recommendation tasks, their suitability for direct numerical forecasting from wearable data has not been systematically studied. This paper presents a unified statistical–semantic aggregation framework that transforms daily wearable-device time series into complementary statistical and structured natural-language representations suitable for text-toregression forecasting with LLMs. We define a one-day-ahead benchmark in which next-day physiological targets are predicted exclusively from aggregated textual summaries, and investigate parameter-efficient fine-tuning using Low-Rank Adaptation (LoRA) for open-weight LLMs. Experiments on multiple public and internal wearable datasets demonstrate that lightweight adaptation consistently improves forecasting accuracy while preserving computational efficiency. In particular, mid-sized open-weight models achieve single-digit heart-rate prediction error and offer a favorable accuracy–cost trade-off compared to substantially larger alternatives. These results indicate that text-based aggregation combined with parameter-efficient adaptation enables practical, scalable, and cost-effective wearable-data forecasting using large language models. |
| Published |
Aachen : CEUR-WS |
| Type |
Conference paper |
| Language |
English |
| Publication date |
2026 |
| CC license |
|
