PaperlayerView exact revision on arXiv

Source-linked AI summary

It's All in the Way You Say It: The Role of Information Representation in LLM-Based Glycemic-Event Prediction

Andrea Apicella, Pasquale Arpaia, Matteo Orefice, Andrea Pollastro, Roberto Prevete

arXiv:2609.08772v1cs.AI

TL;DR

Prompt-based LLM glucose prediction may depend strongly on how physiological information is represented, but evidence for postprandial glycemic-event forecasting remains limited. This study evaluates general-purpose LLMs across representations, information settings, horizons, and prompting regimes, comparing them with supervised models and Gluco-LLM. Supervised models perform best for hyperglycemia, while prompt-based LLMs outperform supervised models for hypoglycemia across the evaluated horizons, with representation effects but no systematic benefit from extra context.

  • Problem

    The study addresses limited evidence about how physiological-information representation and type affect prompt-based postprandial glucose prediction in T1DM.

  • Method

    The study compares supervised models, prompt-based general-purpose LLMs, and adapted LLM forecasting using Raw, Structured, and Narrative representations with CGM-derived or contextual information.

  • Results

    Supervised models achieve the strongest hyperglycemia performance, whereas the best prompt-based configurations outperform supervised results for hypoglycemia at all three horizons.

  • Takeaways & Limitations

    Representation design is a central factor in prompt-based glycemic-event prediction, with no universally optimal representation and no systematic improvement from additional contextual information.

  • Takeaways & Limitations

    Because the evaluation uses a single dataset and LLM pretraining exposure cannot be excluded, the observed behavior may not generalize and possible data leakage cannot be ruled out.

Abstract

from arXiv · show

Large Language Models (LLMs) are increasingly being investigated for physiological time-series prediction, yet their effectiveness may depend not only on the model itself, but also on how physiological information is represented and presented at inference time. This study investigates prompt-based general-purpose LLMs for postprandial hyperglycemia and hypoglycemia prediction in individuals with type 1 diabetes. Using the OhioT1DM dataset, we evaluate multiple open-weight LLMs under zero-shot and few-shot inference across prediction horizons of 30, 60, and 90 minutes. The analysis varies both the textual representation of the available physiological information and the amount of information exposed to the model, ranging from glucose observations alone to derived descriptors and additional contextual variables related to insulin, meals, carbohydrates, and physical activity. Performance is compared with conventional patient-specific supervised models and with Gluco-LLM, a language-model-based architecture explicitly adapted to glucose time-series forecasting. Results show a marked task-dependent behavior. Conventional supervised models achieve the strongest performance for hyperglycemia prediction, whereas the best observed prompt-based LLM configurations improve performance for hypoglycemia across all investigated horizons. The effectiveness of prompt-based inference is also strongly influenced by how physiological information is represented, while providing additional contextual information does not lead to a systematic improvement. Overall, these findings highlight physiological information representation as a central design factor in prompt-based LLM approaches to glycemic-event prediction.

1. Introduction

Postprandial glycemic-event prediction is clinically relevant because glucose dynamics reflect interacting physiological factors. This study evaluates whether prompt-based general-purpose LLMs can predict these events and how representation choices affect performance.

  • Motivation: Postprandial prediction targets hyperglycemic and hypoglycemic events after meals, when carbohydrate absorption, insulin, and individual responses interact.Accurate forecasts may support earlier identification of adverse glucose excursions and more timely preventive interventions.
  • Motivation: Prompt-based LLMs require physiological observations to be transformed into representations compatible with natural-language interaction.Candidate interfaces include direct glucose sequences, clinically meaningful descriptors, and natural-language statements.
  • Research gap: Existing work mainly adapts pretrained LLMs for time-series forecasting or directly queries them, leaving representation effects in prompt-based glucose prediction insufficiently understood.The study focuses this question on postprandial glucose forecasting in T1DM.
  • Study design: The study uses OhioT1DM multimodal data containing CGM, insulin, meals, carbohydrates, behavioral events, and wearable physiological signals.These inputs support analysis of both information representation and the amount of physiological context exposed to models.
  • Research questions: Three research questions compare general-purpose LLMs with supervised models, test representation and information effects, and assess comparison with Gluco-LLM.The evaluation includes Raw, Structured, and Narrative representations with CGM-derived or contextual information settings.
  • Main findings: Performance depends on the prediction task: supervised models are strongest for hyperglycemia, while prompt-based LLMs are particularly competitive for hypoglycemia.Structured inputs are most frequently associated with the strongest configurations across LLM families.

2. Related Works

Prior research spans conventional glucose forecasting, LLM-based time-series adaptation, and conversational CGM analysis. However, representation and information-type effects in prompt-based postprandial glucose prediction remain underexplored.

  • Conventional forecasting: Patient-specific supervised models have been used to predict postprandial glucose extrema from historical CGM measurements.This work provides a conventional modeling context for glycemic-event prediction.
  • LLM forecasting: LLMTime established zero-shot time-series forecasting by representing numerical sequences as text and framing prediction as next-token generation.This paradigm motivated broader use of general-purpose LLMs for numerical forecasting.
  • LLM adaptation: DiabLLM adapts language-model-inspired forecasting architectures and pretrained LLM backbones for blood-glucose prediction in T1D.It differs from directly prompting a general-purpose LLM because the backbone is used for numerical time-series forecasting.
  • Conversational CGM analysis: Other studies have examined GPT-4 for transforming CGM time series into clinically meaningful narratives and for conversational CGM analysis.These efforts focus on analysis and summarization rather than the postprandial prediction task addressed here.
  • Research gap: Prior work had not examined how physiological-information representation and type affect prompt-based postprandial glucose prediction in T1DM.The present study addresses this gap by varying both representation and available physiological information.

3. Method

The method compares conventional supervised models, unchanged general-purpose LLMs prompted with alternative physiological representations, and Gluco-LLM for meal-centered glycemic-event prediction. Within prompt-based inference, it varies representation, available contextual information, prediction formulation, and zero-shot versus few-shot prompting.

  • Comparative framework: Three complementary approaches are compared: patient-specific supervised ML, unchanged general-purpose LLM prompting, and Gluco-LLM.The approaches are evaluated on aligned postprandial prediction tasks and horizons.
  • Information representation: Prompt inputs use Raw, Structured, or Narrative representations of physiological information.Raw serializes the CGM sequence; Structured uses key–value descriptors; Narrative expresses descriptors in natural language.
  • Information availability: The CGMderived setting uses CGM-only descriptors, whereas CGMcontext-derived adds insulin, carbohydrate, meal-timing, and physical-activity variables.These settings test whether prediction benefits from CGM-derived characteristics alone or complementary contextual information.
  • Prediction setting: The meal-centered protocol defines a four-hour postprandial interval, using the first two hours for hyperglycemia and the following two for hypoglycemia prediction.Prediction samples are generated within each meal’s postprandial episode.
  • Prediction formulation: The task supports direct binary classification or regression followed by thresholding of predicted future glucose extremes.Hyperglycemia uses the future maximum, while hypoglycemia uses the future minimum; classification labels require at least two consecutive threshold-crossing CGM readings.
  • Inference protocol: Prompts are generated deterministically and evaluated under zero-shot and few-shot inference conditions.Few-shot prompts include labeled examples using the same representation, information setting, task, and horizon as the target sample.

4. Experimental assessment

The experimental assessment compares patient-specific supervised models, prompt-based general-purpose LLMs, and Gluco-LLM under a common postprandial prediction framework. It evaluates multiple horizons, information configurations, and inference settings on OhioT1DM data.

  • Dataset and preprocessing: The OhioT1DM experiments use patient-specific physiological data, with 11 patients retained after excluding subject 567 for unavailable test-partition meal information.The dataset includes CGM, insulin, meal carbohydrates, fingerstick glucose, and physical activity data.
  • Dataset and preprocessing: Each prediction instance contains 60 minutes of CGM history, and horizons are extended to 30, 60, and 90 minutes.The historical window contains T = 12 measurements on the five-minute sampling grid.
  • Compared approaches: The study evaluates three complementary approaches: conventional supervised learning, unchanged general-purpose LLMs queried through prompts, and the time-series-adapted Gluco-LLM.The supervised classification baselines include Logistic Regression, XGBoost, CNN1D, and LSTM; Gluco-LLM is evaluated in regression.
  • Compared approaches: Prompt-based experiments compare four open-weight instruction-tuned LLM families under zero-shot and few-shot inference without task-specific parameter updates.All LLMs use a 4096-token context window and are executed locally.
  • Research questions: The study examines whether physiological information representation and availability influence prompt-based prediction, alongside comparisons with supervised models and Gluco-LLM.Representations and contextual inputs are varied to address RQ2, while regression comparisons address RQ3.

5. Results

Results show task-dependent differences between supervised models and prompt-based LLMs, while representation strongly affects LLM performance. Additional contextual information does not consistently improve predictions, and the task-specific regression model remains strongest overall.

  • RQ1: General purpose LLMs vs. supervised learning: Supervised models achieve the highest MCC for hyperglycemia at all three prediction horizons.The performance gap narrows as the horizon increases, but the best supervised configuration remains highest throughout.
  • RQ1: General purpose LLMs vs. supervised learning: Prompt-based LLMs achieve higher observed MCC than supervised models for hypoglycemia at every horizon, including zero-shot inference.The advantage is not restricted to few-shot prompting, although the underlying mechanism requires further investigation.
  • RQ2: Information representation and availability: Zero-shot results show no general preference for Raw CGM trajectories: Structured representation is best in most hyperglycemia cases, while hypoglycemia results are more heterogeneous.The comparison concerns Raw, Structured, and Narrative representations under CGM-derived information.
  • RQ2: Information representation and availability: Few-shot results favor Raw representation in most model–horizon configurations, while Structured or Narrative representations still produce the best overall task-level results in several settings.Derived representations are strongest for hyperglycemia at 30 minutes and hypoglycemia at all three horizons when selecting the best model.
  • RQ2: Information representation and availability: Adding insulin, carbohydrate, meal, basal-rate, and exercise information does not systematically improve LLM performance through the evaluated prompt representations.This finding does not establish that those variables are predictively irrelevant.
  • RQ3: Prompt-based inference vs. adapted LLMs: The Cui et al. task-specific model achieves the highest MCC in 4 of 6 regression task–horizon combinations and the lowest RMSE in all six.At 90 minutes, the reported MCC values are 0.234 and 0.436, respectively, for the compared configurations.
  • RQ3: Prompt-based inference vs. adapted LLMs: None of the investigated LLM approaches consistently matches the Cui et al. regression model, which remains the strongest and most consistent approach in the evaluated setting.Gluco-LLM is generally closer to the task-specific model than prompt-based LLMs, but does not consistently surpass it.

6. Discussion

General-purpose LLMs are competitive for hypoglycemia but not hyperglycemia, and their effectiveness depends strongly on representation, prompting, task, and horizon. No representation or prompting strategy is universally best, while explicit time-series adaptation and supervised models retain important advantages.

  • Prompt-based LLMs outperform supervised models for hypoglycemia at all three horizons, while supervised models achieve the highest MCC for hyperglycemia at every horizon.The comparison uses the best observed configurations across evaluated models and information settings.
  • Performance varies across LLM families, prompting strategies, representations, target events, and prediction horizons.The study therefore characterizes prompt-based inference as configuration-dependent rather than uniformly reliable.
  • Structured representations produce the best overall results in most task–prompting–horizon combinations, whereas Raw and Narrative remain preferable in specific settings.Zero-shot results favor Structured in most hyperglycemia cases, while few-shot comparisons shift the advantage toward Raw in most model–horizon configurations.
  • Adding physiologically relevant context does not ensure better prompt-based inference; the effect of insulin, carbohydrate, basal-rate, meal, and exercise information depends on its representation.How variables are represented and integrated into the prompt can matter at least as much as which variables are provided.
  • Few-shot examples improve the best observed result in several cases but increase MCC in only a limited number of matched comparisons.Their effectiveness depends on the model, representation, information setting, task, and prediction horizon.
  • Gluco-LLM generally outperforms prompt-based general-purpose LLMs in regression, but the Cui et al. model remains most reliable, achieving the lowest RMSE in every task–horizon combination.The isolated Gluco-LLM MCC advantage for hypoglycemia at 60 minutes is small relative to run variability.

7. Limitations and future work

The study’s limitations constrain the generality of its conclusions across datasets, model classes, inference variability, and physiological information representations. Future work should broaden evaluation and characterize these factors more systematically.

  • Dataset scope: Experiments use only the OhioT1DM dataset, so observed behavior may not generalize across populations, devices, treatment strategies, or collection protocols.The authors recommend reproducing the framework on additional CGM datasets and assessing cross-dataset robustness.
  • Model scope: The evaluated prompt-based models are open-weight LLMs in the 7-9B parameter range, excluding larger open-weight and proprietary frontier models.Broader scales and architectures are needed to distinguish model-capacity effects from intrinsic limitations of prompt-based physiological reasoning.
  • Evaluation reporting: Table 9 compares regression-based approaches using best observed prompt-based results and ten-seed mean (standard deviation) results for Cui and Gluco-LLM, with lower RMSE preferred.Bold values mark the best overall result for each task, horizon, and metric.
  • Inference variability: Prompt-based configurations were evaluated in a single fixed inference run, leaving sensitivity to demonstrations, decoding settings, and other inference-time variability unquantified.Future studies should repeat experiments across decoding configurations, including temperature settings, and report uncertainty estimates.
  • Representation scope: The investigated representations cover only a limited subset of multimodal physiological inputs, so the lack of systematic contextual benefit is representation-specific rather than evidence of intrinsic irrelevance.The authors call for adaptation strategies and ablations examining heterogeneous signals and in-context demonstrations.

8. Conclusion

The study compares supervised, prompt-based, and adapted language-model approaches for patient-specific postprandial glycemic-event prediction. Results show that performance depends on the event, representation, prompting configuration, and model adaptation, with no uniformly dominant paradigm.

  • Study scope: The evaluation compares conventional supervised ML, prompt-based general-purpose LLMs, and time-series-adapted language models under patient-specific, temporally causal protocols.It covers direct event classification and regression of future glucose extrema.
  • Cross-paradigm findings: Few-shot prompting is configuration-dependent rather than uniformly beneficial, and no uniform ranking separates supervised ML from LLM-based inference.General-purpose LLMs are therefore not intrinsically unsuitable for glucose prediction, but their effectiveness depends on the target event and experimental configuration.
  • Information representation: Physiological information representation substantially affects prompt-based prediction, without a universally optimal representation across LLM families, prompting strategies, events, and horizons.Structured inputs lead when selecting the best result across models in most combinations, while Raw is more frequent under matched few-shot configurations and Narrative is competitive in some hypoglycemia settings.
  • Contextual information: Adding insulin, carbohydrate, meal, basal-rate, and exercise information to CGM-derived inputs does not systematically improve performance.Benefits vary by configuration and representation, indicating that additional information must be effectively integrated through the input interface.
  • Model adaptation: Explicit time-series adaptation is more effective than prompt-only inference for using a pretrained language-model backbone, but the task-specific recurrent model remains strongest and most consistent for regression.Gluco-LLM is generally more competitive than prompt-only inference, particularly for hypoglycemia, while the recurrent baseline has lower RMSE in every evaluated condition and the highest MCC in most comparisons.
  • Overall conclusion: The overall trade-off is between task-specific models’ consistency and prompt-based inference’s training-free adaptability and representational flexibility.Specialized models remain most reliable in several conditions, particularly hyperglycemia classification and continuous glucose prediction, while selected general-purpose configurations can be competitive or superior for hypoglycemia detection.
Loading 2609.08772v1…