Advanced Biostatistical Machine Learning for Complex Biomedical Data: Causal Dose-Response Curve Estimation, ICU EHR Bias and Missing Data Evaluation, and Cancer-Drug Biclustering and Response Prediction - Dissertation

Aug 4, 2026

Upload by Riley

Large-scale biomedical data in critical care, clinical trials, and precision medicine requires robust machine learning methods for complex, high-dimensional settings. This dissertation proposes biostatistical machine learning approaches addressing three challenges: causal dose-response estimation for continuous treatments, evaluation of bias and missing-data patterns in ICU Electronic Health Records, and learning cancer-drug interactions via biclustering and response prediction. Results introduce a Highly Adaptive Lasso (HAL)-based plug-in estimator yielding valid inference without parametric assumptions, demonstrate demographic-linked measurement biases with predictive relevance, and present ImpaCluster for sensitive biclusters and accurate drug-response prediction on unseen cell lines and compounds.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages70

Reading Time2 h 56 min

AccessFree to download

Introduction

HAL-Based Plugin Estimation of the Causal Dose-Response Curve

Challenges in Estimating Causal Dose-Response

HAL-Based Plugin Estimator

Simulation Design

Estimator Evaluations through Simulations

Discussion

Conclusions and Future Work

Measurement Patterns and Implicit Bias in ICU Data

Overlooked Measurement Patterns in ICU EHRs

Measurement Patterns and Statistical Analysis Methods

Implicit Bias and Predictive Power

Discussion

Conclusions and Future Work

Integrative Deep Multi-Learning for Predicting and Biclustering Cancer Drug Responses (ImpaCluster)

Motivation behind Biclustering and Prediction

Architecture of ImpaCluster and Data Applied

Validations of ImpaCluster

Discussion

Conclusion

Concluding Remarks and Future Prospects

Bibliography

UC Berkeley UC Berkeley Electronic Theses and Dissertations Title Advanced Biostatistical Machine Learning for Complex Biomedical Data: Causal DoseResponse Curve Estimation, ICU EHR Bias and Missing Data Evaluation, and Cancer-Drug Biclustering and Response Prediction Permalink [https://escholarship.org/uc/item/41g2m9rz](https://escholarship.org/uc/item/41g2m9rz) Author Shi, Junming Publication Date 2024 Peer reviewed|Thesis/dissertation [eScholarship.org](eScholarship.org) Powered by the California Digital Library University of California Advanced Biostatistical Machine Learning for Complex Biomedical Data: Causal Dose-Response Curve Estimation, ICU EHR Bias and Missing Data Evaluation, and Cancer-Drug Biclustering and Response Prediction by Junming Shi A dissertation submitted in partial satisfaction of the requirements for the degree of Doctor of Philosophy in Biostatistics and the Designated Emphasis in Computational and Genomic Biology in the Graduate Division of the University of California, Berkeley Committee in charge: Professor Alan Hubbard, Co-chair Professor Haiyan Huang, Co-chair Professor Mark van der Laan Fall 2024 Advanced Biostatistical Machine Learning for Complex Biomedical Data: Causal Dose-Response Curve Estimation, ICU EHR Bias and Missing Data Evaluation, and Cancer-Drug Biclustering and Response Prediction Copyright 2024 by Junming Shi 1 Abstract Advanced Biostatistical Machine Learning for Complex Biomedical Data: Causal Dose-Response Curve Estimation, ICU EHR Bias and Missing Data Evaluation, and Cancer-Drug Biclustering and Response Prediction by Junming Shi Doctor of Philosophy in Biostatistics and the Designated Emphasis in Computational and Genomic Biology University of California, Berkeley Professor Alan Hubbard, Co-chair Professor Haiyan Huang, Co-chair The advent of large-scale biomedical data in fields such as critical care, clinical trials, and precision medicine has driven the need for robust machine learning methodologies capable of handling complex, high-dimensional datasets. This dissertation introduces a series of innovative biostatistical machine learning approaches designed to address three critical challenges in biomedical research: (1) estimating causal dose-response curves for continuous treatments, (2) identifying and addressing biases and missing data patterns in Electronic Health Records (EHRs), and (3) learning cancer-drug interactions through biclustering and prediction in precision oncology. The first project introduces, implements, and evaluates the Highly Adaptive Lasso (HAL)-based plug-in estimator for estimating causal dose-response curves without relying on parametric assumptions that risk model misspecification. Based on theoretical proofs by van der Laan and through extensive simulations, the HAL estimator provides valid inference with robust confidence intervals for continuous dose-response relationships, outperforming traditional methods. The second project develops a framework for analyzing measurement patterns in ICU EHRs. This analysis reveals systematic biases linked to demographic factors such as race, sex, and age. In addition, these measurement patterns demonstrate strong predictive power for patient outcomes. The findings underscore the need for fairer and more accurate machine learning models in healthcare. The final project presents Integrative Deep 2 Multi-Learning for Predicting and Biclustering Cancer Drug Responses (ImpaCluster), a deep multi-task learning algorithm that integrates cancer omics data, drug molecular profiles, and drug response data. ImpaCluster identifies sensitive cancer-drug biclusters with shared molecular features and accurately predicts drug responses, even for unseen cancer cell lines and compounds, providing a powerful tool for advancing personalized cancer therapies. These contributions address significant gaps in biostatistical an

FAQs

What three biomedical research challenges does the dissertation address?

It addresses causal dose-response estimation for continuous treatments, bias and missing-data evaluation in ICU Electronic Health Records, and cancer-drug interaction learning via biclustering and response prediction.

How does the dissertation estimate causal dose-response curves?

It introduces a Highly Adaptive Lasso (HAL)-based plug-in estimator that avoids parametric assumptions and uses theoretical results to provide valid inference with robust confidence intervals.

What does ImpaCluster contribute to cancer therapy prediction?

ImpaCluster performs deep multi-task learning to identify sensitive cancer-drug biclusters with shared molecular features and to predict drug responses, including for unseen cancer cell lines and compounds.

Aug 4, 2026

Upload by Riley