Research map / Architectures and efficiency / RNNs and normalization
Robust and Adaptive AI Models for Medication Usage Forecasting Using ICD-9/10 Code (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds an unsupervised NLP/PCA-based vector representation of daily ICD-9/10 codes and medication usage from EHR data, then ensembles TFT and AR-LSTM forecasters with Bayesian changepoint-triggered finetuning to forecast 10-day medication usage.
Problem
Accurate forecasting of medication usage and ICD-9/10 code streams is needed for healthcare logistics, especially during high-demand periods such as pandemics, outbreaks, wartime, or disasters.
Approach
Combine incremental PCA and TF-IDF vector encodings of ICD codes and FDA medication lists with an ensemble of TFT and AR-LSTM models trained at multiple lookback horizons, plus Bayesian changepoint detection to trigger model updates when forecasting error drifts.
Key results
- Ensemble of TFT and AR-LSTM achieves average forecasting error of 5.2% at day 5.
- Average forecasting error rises to 18.1% at day 10, with some medications exceeding 50% error.
- In the day-5 table, the ensemble is closest to actual usage for several medications, e.g., Prednisone 23/25 vs VAR 18/25.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the median ensemble of TFT and AR-LSTM models trained with different lookback horizons | Medication forecasting framework · MIMIC-IV ER and MIMIC-III, 5-day forecast | mean error: 5.2% | Helped |
| Use the median ensemble of TFT and AR-LSTM models trained with different lookback horizons | Medication forecasting framework · MIMIC-IV ER and MIMIC-III, 10-day forecast | mean error: 18.1% | Helped |
| Use the median ensemble instead of the individual models | Medication forecasting framework · 5-day medication forecasting across medications in Table 1 | forecasting result (as reported) | Mixed |
| Train TFT and AR-LSTM with Stochastic Weight Averaging (SWA) | TFT and AR-LSTM · Medication forecasting | forecasting performance | Helped |
Surprising findings
- Maximum 10-day forecasting error exceeds 50% for some medications, showing performance varies considerably across drugs.
Limitations
- Performance varies greatly across medications, with some errors above 50% at day 10.
- Error grows substantially with forecast horizon (5.2% at day 5 vs 18.1% at day 10).
- Single-author student abstract with limited methodological detail and no released code.
Experimental setup
Datasets: MIMIC-IV ER MIMIC-III FDA full medication list
Benchmarks: MIMIC-IV ER MIMIC-III
Metrics: forecasting error (%) dynamic time warping distance MAE RMSE
Baselines: Linear Drift Model Vector Autoregression (VAR)
Builds on: Temporal Fusion Transformer Incremental PCA Bayesian online changepoint detection Stochastic Weight Averaging Dynamic time warping
Closest papers by meaning
- MIRA: Medical Time Series Foundation Model for Real-World Health Data (NeurIPS 2025 · similarity 0.83)
- Context Clues: Evaluating Long Context Models for Clinical Prediction Tasks on EHR Data (ICLR 2025 · similarity 0.81)
- Trustworthy and Practical AI for Healthcare: A Guided Deferral System with Large Language Models (AAAI 2025 · similarity 0.81)
- One Loss to Rule Them All: Marked Time-to-Event for Structured EHR Foundation Models (NeurIPS 2026 · similarity 0.80)
- Risk Horizons: Structured Hypothesis Spaces for Longitudinal Clinical Prediction (NeurIPS 2026 · similarity 0.80)
- MedChain: Bridging the Gap Between LLM Agents and Clinical Practice with Interactive Sequence (NeurIPS 2025 · similarity 0.80)
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.