Research map / Time series / Foundation models for time series
Can we generate portable representations for clinical time series data using LLMs?
Zongliang Ji, Yifei Sun, Andre Carlos Kajdacsy-Balla Amaral, Anna Goldenberg, Rahul G Krishnan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace grid imputation/TSDE/TimesFM inputs with LLM summarize-then-embed (Record2Vec) | ICU downstream predictors · HiRID, MIMIC-IV, PPICU in-distribution | task metrics: best on 13 of 15 tasks | Improved |
| Use Record2Vec instead of numeric baselines for cross-site transfer | HiRID/MIMIC-IV -> PPICU · Cross-site transfer | task metrics: best in 10 of 10 columns (2 ties with TimesFM) | Improved |
| Add LLM summarization before embedding (vs embedding raw serialization) (ablation) | Record2Vec · Cross-site transfer (30 tasks) | rank: no-summary concentrated in worst ranks; ~25x fewer tokens | Improved |
| Add LLM summarization before embedding (vs raw serialization) (ablation) | Record2Vec · In-distribution (15 tasks) | rank | No clear effect |
| Use Llama-3.1 instead of Gemini-2.0 Flash or MedGemma as summarizer (ablation) | Record2Vec · In-distribution and cross-site | rank | Hurt |
| Vary prompt style (zero-shot, CoT, trend, ICD) (ablation) | Record2Vec with Gemini-2.0 Flash · In-distribution and cross-site | rank: slight edge for ICD prompts in transfer | No clear effect |
| Replace LLM summary with fixed template (Record2Vec Template) (ablation) | Record2Vec · HiRID -> PPICU mortality AUROC | AUROC: 0.72 -> 0.72 | No clear effect |
| Replace LLM summary with fixed template (Record2Vec Template) (ablation) | Record2Vec · MIMIC-IV -> PPICU length of stay | MAE: 0.49 -> 0.77 | Hurt |
| Use Record2Vec embeddings rather than baselines for demographic probes | Record2Vec · Age/sex recoverability | age MAE, gender | No clear effect |
| Few-shot fine-tune Record2Vec with 16 target samples | Record2Vec · PPICU few-shot mortality | AUROC | Improved |
| Use Record2Vec instead of TSDE on PPICU mortality in-distribution | Record2Vec · PPICU in-distribution | Mortality AUROC: 0.890 vs 0.890 | No clear effect |
Closest papers by meaning
- MedJourney: Benchmark and Evaluation of Large Language Models over Patient Clinical Journey (NeurIPS 2024 · similarity 0.87)
- TEST: Text Prototype Aligned Embedding to Activate LLM's Ability for Time Series (ICLR 2024 · similarity 0.85)
- Integrating Sequence and Image Modeling in Irregular Medical Time Series Through Self-Supervised Learning (AAAI 2025 · similarity 0.85)
- LLMs can construct powerful representations and streamline sample-efficient supervised learning (NeurIPS 2026 · similarity 0.85)
- Context Clues: Evaluating Long Context Models for Clinical Prediction Tasks on EHR Data (ICLR 2025 · similarity 0.85)
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
