PocketTSLM: A Mobile Time-Series Model for Privacy Protection

Paper: PocketTSLM: Deploying an On-Device Soft-Prompt Time-Series Language Model for Privacy-Preserving Smartphone Reasoning

Authors: Max Rosenblattl, Nikhil Krishnaswamy, Leon Nissen, Manuel Milling, Nicolas Zumarraga, Ning Wang, Thomas Kaar, Robert Jakob, Bj"orn W. Schuller, Elgar Fleisch, Patrick Langer, Paul Schmiedmayer

Executive Summary

The research introduces PocketTSLM, a mobile-native deployment of a time-series language model designed to handle physiological data on consumer smartphones effectively. Traditionally, mobile systems tend to treat time-series data — such as those collected by consumer phones and wearables — as static summaries or text. This approach fails to utilize the temporal aspects of the data, which are crucial for personal health applications needing to analyze, retrieve, or summarize sensor data in context.

The study highlights the limitations of cloud-hosted models, which transfer sensitive health data off the device, resulting in potential privacy concerns and increased latency along with higher operating costs for each query. Existing mobile alternatives that convert time-series data into statistical features or serialize them into text often compromise the data's temporal structure or exceed the device's memory and processing capabilities.

PocketTSLM addresses these challenges by processing time-series data directly on the smartphone. Instead of converting time-series data into text, it uses encoded time-series windows as embeddings, which preserves the temporal structure and enhances processing efficiency. The implementation employs a numerical parity protocol to ensure that the token sequences generated by PocketTSLM on a smartphone match those produced by a Python reference model, confirming the accuracy of the mobile version.

Benchmarks reveal that PocketTSLM significantly outperforms traditional text-serialized approaches. It uses approximately 12 times fewer time-series tokens, resulting in lower prompt-prefill costs and allowing for more efficient processing of longer time-series windows. The model has demonstrated its capability by successfully processing 30-second clips of electroencephalography (EEG) data for sleep staging and 1,000-sample (10 seconds) electrocardiogram (ECG) data, both of which are critical for health monitoring applications.

This research matters because it represents a step towards more effective and privacy-conscious health monitoring solutions. By enabling on-device processing, PocketTSLM could allow individuals to analyze their health data without the need to upload it to the cloud, thus reducing privacy risks and latency. The results, drawn from benchmarks rather than real-world applications, suggest that this approach may be particularly beneficial for health apps that rely on real-time data analysis and timely feedback to users.

Academic Abstract

Consumer phones and wearables increasingly collect physiological time-series data, yet most mobile systems treat these streams as summaries or text rather than as temporal evidence. This representation is poorly matched to personal health systems that need to retrieve, route, or summarize sensor evidence while preserving temporal structure. Cloud-hosted time-series language models avoid smartphone limits, but they repeatedly move sensitive physiology off the device and add latency and per-query compute cost. Mobile model alternatives that aggregate time-series data into statistical features or serialize raw samples as text either compress temporal structure or exhaust the context and memory budgets of small on-device models. We present PocketTSLM, a mobile-native deployment of OpenTSLM SoftPrompt, the soft-prompt variant of the OpenTSLM time-series language model. PocketTSLM runs fully on a consumer smartphone by injecting encoded time-series windows as embeddings instead of tokenizing time series as text. A numerical parity protocol verifies identical greedy-decoded token sequences between the Swift port and a Python reference using Apple's MLX machine-learning framework. Answers also match the PyTorch reference, although rationale wording can differ. An on-device context-economics benchmark compares the soft-prompt path with a text-serialized baseline on the same backbone. In our benchmark format, soft-prompting uses approximately 12 times fewer time-series tokens and reduces prompt-prefill cost. PocketTSLM processes 30-second Sleep-EDF electroencephalography clips for sleep staging and a 1,000-sample (10 s) 12-lead electrocardiogram clip from ECG-QA. In comparison, text-serialized baselines exceed the phone's context or memory budget at longer time-series windows. All inference runs entirely on the device, with no health data transmitted during inference.