Acute Kidney Injury (AKI) affects 13% to 18% of hospitalized patients in the U.S., with men at significantly higher risk than women. The mechanisms behind sex disparities remain poorly understood. This secondary analysis utilizes the MIMIC-III and MIMIC-IV data to investigate sex-specific predictive models for AKI outcomes. We will develop deep learning (DL) models that incorporate sex differences using multi-modal data, including demographics, clinical, and text predictors. We will compare our model's performance against traditional regression and other DL models. Aim 2 seeks to identify sex-specific risk factors for AKI, Chronic Kidney Disease, and mortality. We will interpret the sex-exposure interactions through novel metrics. By identifying and simulating modifiable risk factors, our models can provide assessment of personalized risks. Our findings will improve understanding of AKI and inform targeted clinical interventions. Results and code will be shared openly.
This secondary analysis aims to investigate sex differences in Acute Kidney Injury (AKI) and its progression to Chronic Kidney Disease (CKD) using the MIMIC-III and MIMIC-IV databases. AKI is a critical condition, affecting approximately 13% to 18% of hospitalized patients in the U.S., with higher incidence in men compared to women. Furthermore, men are more likely to experience severe types of AKI, associated complications, and death. Despite significant disparities, the underlying mechanisms remain poorly characterized, calling for further investigations.
The primary data source for this project will be the Medical Information Mart for Intensive Care (MIMIC)-III and MIMIC-IV data, which contain a wealth of clinical data, including demographics, laboratory values, comorbidities, and treatment interventions. MIMIC-III comprises over 60,000 hospital admissions, while MIMIC-IV expands this with additional records and updated data structures. The datasets provide structured and unstructured data, such as free-text clinical notes, enabling a comprehensive analysis of factors influencing AKI.
Aim 1: We will develop deep learning (DL) prediction models to identify AKI and its outcomes while incorporating sex-specific effects. Our proposed transformer model will utilize multi-modal variables from MIMIC databases, accommodating correlated numerical variables, longitudinal data, and free-text data through an attention mechanism. We will evaluate our model's performance against traditional methods (regression, CART, Random Forest, LightGBM, XGBoost) using metrics such as AUC, accuracy, precision, recall, and F1 score.
Aim 2: We will investigate common and unique risk factors for AKI, CKD, and mortality by exploring interactions between predictors and sex. We will develop methods to quantify the contributions of individual predictors and their interactions, using a novel metric for two-way interactions.
We will be able to identify fixed and modifiable features for AKI risk and outcomes. Using sex-specific modifiable exposures, we will be able to simulate scenarios that predict risk changes based on varying conditions (e.g., blood pressure, medication choices).
The anticipated timeline is
The findings from this project will yield critical insights into the role of sex in AKI and CKD outcomes, thereby informing targeted clinical interventions. Ultimately, we aim to establish analytical frameworks for investigating health disparities in AKI diseases, contributing to more effective prevention and management strategies tailored to individual patients.
Primary Outcomes and Subject Selection
Predictor Variables
AKI Risk Prediction Model
An improved transformer model will be developed to handle longitudinal and multi-modal data. The attention mechanism will capture complex relationships among variables, facilitating efficient processing.
CKD and Death Outcome Prediction Model
A similar model setup will be used for predicting CKD within five years post-AKI and for time-to-event outcomes like death.
Interaction Explainability
Two-way partial dependence plots will elucidate interactions between sex and other predictors. The H-statistic will quantify the contribution of interactions, enhancing understanding of sex disparities in health outcomes.
Key Research Findings and Expected Outcomes
Intermediate variables and resulted individual predictions from multiple models will be deposited in DataVerse, for future research. Python programs will be shared on GitHub, allowing others to replicate analyses and adapt methods for other diseases with subgroup differences. Datasets and code will include detailed methodologies and usage instructions. We plan to write manuscripts based on our findings for peer-reviewed journals.
Acute Kidney Injury (AKI) affects 13% to 18% of hospitalized patients in the U.S., with men at a significantly higher risk. Contributing factors include variations in baseline kidney function, cardiovascular diseases, hypertension, and hormonal influences. Although standard AKI management strategies like hydration and blood pressure control benefit both sexes, sex-specific responses to medications, such as RAAS inhibitors, indicate a need for tailored treatment protocols. Men often experience a more aggressive decline in renal function and are more likely to progress to Chronic Kidney Disease (CKD) and end-stage renal disease (ESRD) compared to women. Women, particularly pre-menopausal woman, generally have better recovery rates.
Sex differences in AKI can be partly attributed to hormonal, immune, and hemodynamic factors. This research aims to identify a broad range of sex-specific risk factors and treatment responses to guide personalized prevention and treatment strategies. Current AKI management applies uniformly across sexes, despite observable disparities. By elucidating these differences, clinicians can tailor treatment plans to account for physiological and pharmacological variances.
The project will advance knowledge in nephrology, sex health disparities, and predictive modeling. By developing advanced deep learning models with extensive MIMIC-III and MIMIC-IV data, this research will contribute to data science and biomedical research. It aims to address the limitations of existing prediction models by offering innovative frameworks that tackle correlated data more efficiently, account for dependency and trends in longitudinal data, and quantify complex interactions in high-dimensional health data. The incorporation of multi-modal data and sophisticated models will improve predictive accuracy, allowing for timely interventions and better clinical resource allocation.
A critical aspect of this research is the method development to interpret two-way interactions within deep learning models. Understanding how individual predictors interact—especially in the context of sex—can identify key factors and pathways influencing health outcomes. Current techniques for model interpretation primarily focus on feature importance but neglect interaction effects, which are essential for informed decision-making in clinical practice. By addressing this gap, the project will not only enhance the interpretability of DL models but also contribute to a more nuanced understanding of health disparities.
Ultimately, this research has the potential to improve patient outcomes by fostering personalized healthcare strategies, thus advancing the understanding of sex differences in AKI and informing targeted interventions in nephrology.
Dr. Feng has 20 years of developing statistical models in big data and complicated study designs. She developed genetic models in testing familial segregation, mapping genes, adjusting for sampling bias, haplotype-based association, parent-of-origin effects, joint calling of copy number variations and SNPs, high dimensional microbiome, and Mendelian Randomization. Her expertise focuses on incorporating complicated relationships among subjects or predictors to improve the accuracy, power, and prediction of outcomes in emerging models . Her recent efforts includes improving transformer models for correlated and longitudinal predictors. Dr. Feng has collaborated on genetic projects in AKI or related outcomes.
With over a decade of experience in machine learning and data science, Dr. Lass has led the development of predictive models across industries, integrating complex systems and large-scale data. She spearheaded generative AI and NLP initiatives, developing strategies for LLMs (Language Language Models), experimenting with RAG pipelines, and advancing recommendation systems. Her expertise spans bias detection frameworks, exploration-exploitation strategies, and continuous model deployment using MLOps, with a focus on delivering impactful, cross-domain AI solutions.
Dr. Feng and Dr. Lass will work closely to obtain data, design models, implement analyses, interpret results, and write reports. They will meet at least twice a week for progress updates and problem-solving.
Data Preprocessing and Feature Engineering
Text data will be preprocessed for consistency. Named Entity Recognition (NER) using BioBERT will standardize extracted entities. Categorical variables will undergo one-hot encoding, while rare categories will be merged.
Model Training and Testing
We will use 10-fold cross-validation to avoid overfitting. The models will be developed using MIMIC-III, and tested using MIMIC-IV. Classification model performance will be evaluated using accuracy, precision, recall, F1-Score, and AUC-ROC. Survival model performance will be evaluated using mean absolute error (MAE), C-index, and AUC-ROC.
Missing Data
For CKD and mortality, we will expect small percentage of missing data due to loss to follow-up or incomplete data, and will conduct sensitivity analysis using the inverse-probability-weighting method, assuming “missing at random". We will estimate the probability of having an outcome using a DL model.
