[Paper Review] Blood lactate concentration prediction in critical care patients: handling missing values

Behrooz Mamandipoor, Mahshid Majd+2 more|arXiv (Cornell University)|Oct 3, 2019

Machine Learning in Healthcare36 references4 citations

TL;DR

This study proposes a machine learning framework to predict blood lactate levels in ICU patients using clinical time series data, focusing on handling missing values via imputation strategies. It demonstrates that LSTM models with missing value indicators achieve a mean absolute error of 0.665, outperforming mean-based imputation and establishing a clinically relevant benchmark for lactate prediction.

ABSTRACT

Blood lactate concentration is a strong indicator of mortality risk in critically ill patients. While frequent lactate measurements are necessary to assess patient's health state, the measurement is an invasive procedure that can increase risk of hospital-acquired infections. For this reason we formally define the problem of lactate prediction as a clinically relevant benchmark problem for machine learning community so as to assist clinical decision making in blood lactate testing. Accordingly, we demonstrate the relevant challenges of the problem and its data in addition to the adopted solutions. Also, we evaluate the performance of different prediction algorithms on a large dataset of ICU patients from the multi-centre eICU database. More specifically, we focus on investigating the impact of missing value imputation methods in lactate prediction for each algorithm. The experimental analysis shows promising prediction results that encourages further investigation of this problem.

Motivation & Objective

• To define blood lactate concentration prediction as a clinically relevant machine learning benchmark for critical care.
• To evaluate the impact of various missing value imputation techniques on lactate prediction performance across multiple algorithms.
• To compare traditional machine learning and deep learning models on a large, multi-center ICU dataset with high temporal granularity.
• To establish a reproducible baseline for future research in lactate-guided clinical decision making using EHR data.
• To promote model interpretability and external validation for clinical deployment readiness.

Proposed method

• Formalized lactate prediction as a regression problem: predict lactate at time $ t + \beta $ using clinical measurements from the prior $ \alpha $ hours.
• Used the multi-center eICU database (13,464 patients, 12.2M records) with 2-hour temporal binning for time-series modeling.
• Applied diverse imputation methods: mean imputation, forward fill, and indicator-based imputation (missingness as a feature).
• Evaluated performance using Mean Absolute Error (MAE) across 13,464 ICU patients with 14,477 stays.
• Trained and compared traditional models (e.g., linear regression) and deep learning models (e.g., LSTM) on imputed and raw data.
• Employed missing value indicators as features to capture informative absence, improving model performance.

Experimental results

Research questions

• RQ1 How does missing value imputation impact the performance of machine learning models in predicting blood lactate levels in ICU patients?
• RQ2 Which imputation strategy—mean-based, forward fill, or missingness indicators—yields the best predictive performance for lactate concentration?
• RQ3 Can deep learning models like LSTM outperform traditional models in lactate prediction when handling missing data in ICU time series?
• RQ4 Is the proposed lactate prediction framework a viable and generalizable benchmark for future clinical machine learning research?
• RQ5 How does the inclusion of missingness indicators as features affect model interpretability and predictive accuracy?

Key findings

• The LSTM-based model achieved the lowest Mean Absolute Error (MAE) of 0.665 in predicting blood lactate levels across 13,464 ICU patients.
• Indicator imputation—where missingness is encoded as a binary feature—produced the highest predictive performance, outperforming mean-based and forward-fill imputation.
• The inclusion of missingness indicators significantly improved model performance, indicating that missing values carry clinically informative signals.
• The study establishes a reproducible benchmark for lactate prediction using the eICU database, with publicly available code for replication.
• Lactate prediction using machine learning shows strong potential to reduce invasive testing and support real-time clinical decision making.
• Future work should explore classifying lactate levels into severity categories and applying advanced models like GRU-D to handle informative missingness.