Large language models forecast patient health trajectories enabling digital twins - Scorecard - MDSpire

Large language models forecast patient health trajectories enabling digital twins

  • By

  • Nikita Makarov

  • Maria Bordukova

  • Papichaya Quengdaeng

  • Daniel Garger

  • Raul Rodriguez-Esteban

  • Fabian Schmich

  • Michael P. Menden

  • October 1, 2025

  • 0 min

Share

Clinical Scorecard: Predicting Patient Health Pathways through Large Language Models for Digital Twin Applications

At a Glance

CategoryDetail
ConditionVarious clinical conditions including non-small cell lung cancer, intensive care unit patients, and Alzheimer’s disease
Key MechanismsUse of large language models (LLMs) to create digital twins that forecast longitudinal clinical trajectories from electronic health records without data imputation or normalization
Target PopulationPatients with diverse clinical conditions represented in electronic health records
Care SettingClinical settings including intensive care units, oncology, and neurology care environments

Key Highlights

  • DT-GPT leverages EHR data to predict patient-specific health trajectories overcoming challenges like missingness, noise, and limited sample sizes
  • DT-GPT outperforms state-of-the-art machine learning models in forecasting clinical variables across multiple datasets
  • The model supports zero-shot forecasting and offers explainability through a human-interpretable interface

Guideline-Based Recommendations

Diagnosis

  • Utilize longitudinal EHR data to inform digital twin models for comprehensive patient trajectory prediction
  • Incorporate multivariate clinical variables simultaneously to capture interdependencies in patient data

Management

  • Apply DT-GPT digital twin models to support treatment selection and adverse event mitigation through personalized forecasting
  • Leverage zero-shot prediction capabilities to anticipate clinical variables without prior task-specific training

Monitoring & Follow-up

  • Use digital twins for continuous patient monitoring by simulating real-time personalized responses to interventions
  • Employ explainable AI interfaces to facilitate clinical interpretation of forecasted trajectories

Risks

  • Be aware of data quality issues inherent in EHRs such as heterogeneity, sparsity, and rare events that may affect model performance
  • Consider limitations of existing models that assume channel independence which may not capture biological correlations

Patient & Prescribing Data

Patients represented in EHR datasets including ICU, NSCLC, and Alzheimer’s disease cohorts

Digital twin forecasts can inform personalized treatment decisions and clinical trial design by predicting patient-specific outcomes over time

Clinical Best Practices

  • Integrate diverse data streams including demographics, diagnoses, lab results, and treatments for holistic patient modeling
  • Employ LLM-based digital twins that do not require complex data preprocessing such as imputation or normalization
  • Use models capable of handling missing and noisy real-world data to improve clinical forecasting accuracy
  • Adopt explainable AI tools to enhance clinician trust and facilitate decision support

References

Original Source(s)

Related Content