Temporal Data Requirement for Predicting Unplanned Hospital Readmissions
| Authors | Ramin Mohammadi et al. |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2605.00738 |
| Download | |
| Categories | cs.LG |
Abstract
With the proliferation of Electronic Health Records (EHRs), a critical challenge in building predictive models is determining the optimal historical data time window to maximize accuracy. This study investigates the impact of various observation windows ranging from the day of surgery to three years prior on predicting 30-day readmission following hip and knee arthroplasties. The dataset encompasses both structured encounter records (over 4 million) and unstructured clinical notes (80,000) from 7,174 patients. To extract meaning from the clinical notes, we employed a suite of non neural (BOW, count BOW, TF IDF, LDA) and neural encoders (BERT, 1D CNN, BiLSTM, Average). We subsequently evaluated models utilizing clinical notes alone, structured data alone, and a combination of both modalities. Our results demonstrate that the optimal time window for unstructured clinical notes is significantly shorter than for structured data, maximum predictive performance was achieved using notes from just three to six months prior to surgery. In contrast, performance using structured data improved as the time window lengthened, but strictly plateaued after twelve months. These modality-specific temporal patterns remained consistent regardless of model complexity or encoder type. Ultimately, these findings challenge the general assumption that more historical data inherently yields better machine learning predictions, establishing targeted time-window guidelines for optimizing readmission prediction models.
Engineering Breakdown
Plain English
This paper addresses a fundamental question in healthcare ML: how much historical patient data do you actually need to predict 30-day hospital readmissions after hip and knee surgery? The researchers tested observation windows ranging from surgery day to three years prior, using 4+ million structured records and 80,000 clinical notes from 7,174 patients. They compared multiple text encoding approaches (both traditional methods like TF-IDF and neural methods like BERT) and evaluated models trained on clinical notes alone, structured data alone, or combined. The key finding is that the optimal historical window varies by data modality and model type, but the paper quantifies exactly how much historical context provides diminishing returns—critical for production systems where pulling years of data increases latency and storage costs.
Core Technical Contribution
The paper's core contribution is empirical quantification of the temporal data requirement trade-off in clinical prediction tasks. Rather than assuming 'more history is always better,' the authors systematically measure how prediction accuracy changes as the observation window expands from 1 day to 1,095 days, isolating the effect of temporal depth from other modeling choices. They do this across multiple encoding strategies (non-neural: Bag-of-Words, TF-IDF, LDA; neural: BERT, 1D CNN, BiLSTM, Average pooling) and data modalities, creating a reusable framework for determining minimal sufficient historical context. This is novel because prior readmission prediction work typically uses fixed, arbitrary time windows without quantifying their necessity.
How It Works
The system processes two data streams in parallel: structured EHR encounters and unstructured clinical notes. For each patient, they construct multiple datasets with different observation windows (e.g., last 30 days, 90 days, 365 days, etc., up to 3 years pre-surgery). Clinical notes enter separate preprocessing pipelines: traditional approaches tokenize and apply Bag-of-Words, TF-IDF, or LDA to create fixed-size feature vectors; neural approaches pass raw text to BERT for contextual embeddings, 1D CNNs for local feature extraction, or BiLSTM for sequential processing, followed by average pooling to create patient-level representations. The structured and encoded clinical data are concatenated and fed into a classification model (likely logistic regression or gradient boosting) to predict 30-day readmission. The process repeats for each observation window, and the accuracy curves are compared to identify the point of diminishing returns.
Production Impact
For engineers building readmission prediction systems, this work directly informs data pipeline architecture and operational costs. Instead of defaulting to 'fetch all available history' (which could mean querying 3 years of records per patient with associated I/O latency), you can configure your system to only pull, say, 180 days of data if empirical validation shows performance plateaus there. This reduces database query time, ETL load, and feature engineering complexity. The multimodal evaluation (structured + notes) helps you decide whether investing in NLP infrastructure (BERT models, GPUs for inference) is worth the incremental accuracy gain over simpler structured-data-only models. The paper quantifies this trade-off explicitly, so you can optimize for your latency SLA: if your model only needs 90 days of history for 95% of max accuracy, you save significant compute during both training and inference.
Limitations and When Not to Use This
The study is scoped narrowly to a single procedure type (hip and knee arthroplasty) at what appears to be one health system, so results may not generalize to other surgical populations or non-surgical conditions where disease progression timescales differ. The paper is incomplete in the abstract—it cuts off before reporting actual results, so we cannot assess whether the conclusions hold across different patient demographics, whether seasonal effects matter, or how the optimal window changes by age or comorbidity profile. There's an implicit assumption that past patient data is reliably available and complete, which fails at institutions with fragmented EHR systems or poor note documentation. Finally, the paper doesn't address the temporal validation problem: when training and test sets span different time periods, the optimal window found in historical data may not match what works in live prediction, especially if clinical practices or coding standards have shifted.
Research Context
This work sits within the well-established readmission prediction literature (models predicting 30-day readmission have been extensively studied) but makes a methodological contribution by formalizing the temporal data requirement question. It builds on prior work using deep learning for clinical NLP (BERT for clinical text, BiLSTM for sequential EHR data) and multi-modal fusion in healthcare ML. The research is timely given the explosion of EHR adoption and the practical pain point that health systems often waste compute ingesting decades of history for predictions that plateau much earlier. It opens a research direction around 'minimal sufficient context' for various clinical tasks—the hypothesis that different outcome types (readmission vs. mortality vs. adverse events) likely have different optimal windows.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::