Developing a Machine Learning Tool for Dynamic Cancer Treatment Strategies

Jiaming ZengAAAI 2020 · DOI

Abstract

Modern oncology faces a significant challenge in personalizing treatment sequences for patients. As therapeutic options expand, clinicians require tools that can account for individual patient trajectories and the optimal timing of interventions. This research proposes a decision-support framework that integrates Natural Language Processing (NLP), causal inference, and decision analysis. By utilizing longitudinal Electronic Medical Record (EMR) data, the proposed tool aims to move beyond population-level averages to provide personalized, dynamic treatment recommendations.

Introduction

Cancer treatment is increasingly becoming a "moving target." With the rapid emergence of new chemotherapies, immunotherapies, and surgical techniques, the complexity of clinical decision-making has scaled exponentially. Clinicians must decide not only which combination of therapies to use but also the precise timing for switching treatments based on a patient's unique response. Current medical practices often rely on population-based studies that lack the granular detail necessary for truly personalized care. This research addresses the critical need for tools that can navigate longitudinal confounding and provide actionable insights in a dynamic clinical setting.

Methodology

The researcher proposes a three-step interdisciplinary approach to transform raw clinical data into a robust decision-support system.

1. Data Structuring via NLP

A vast amount of critical clinical information is trapped in unstructured formats, such as nursing notes and radiology reports. To unlock this data, the research employs a hybrid NLP pipeline. Building on established semantic data mining techniques and neural embeddings, this process converts free-text clinical notes into "computable" vector-based datasets, allowing high-dimensional clinical details to be used in statistical modeling.

2. Causal Inference with Counterfactual Gaussian Processes (CGP)

Traditional models like Markov Decision Processes (MDPs) often struggle with the irregular time intervals common in real-world medical visits. To overcome this, the author adopts the Counterfactual Gaussian Process (CGP) framework. This approach allows for causal inference on longitudinal data, enabling the model to predict how a patient’s disease might progress under various hypothetical treatment sequences. This is essential for understanding the "counterfactual"—what would have happened if a different treatment path had been taken.

3. Integration of Decision Analysis

The final component of the framework bridges causal inference (understanding the effect of a treatment) with decision analysis (determining the optimal action). By combining these fields, the tool aims to provide recommendations that account for patient-specific characteristics, potential side effects, and individual preferences, creating a comprehensive framework for clinical AI.

Data and Validation

The research utilizes the Stanford Cancer Institute Research Database (SCIRDB), a massive dataset containing records for over 100,000 patients since 2008. This database integrates information from EPIC, STARR, surgical pathology, and radiation oncology, providing a rich foundation of inpatient stays, radiology, and lab results.

As this work was presented as a research plan, the primary focus is on a rigorous validation strategy:

  • RCT Reproduction: The model will be validated by its ability to reproduce the results of established, high-quality randomized clinical trials (RCTs) for localized prostate, esophageal, and oropharynx cancers.
  • Benchmarking: The CGP model will be benchmarked against existing MDP models to demonstrate its superior ability to handle the "messy" temporal nature of observational EMR data.

Discussion and Future Work

While the proposed framework offers a sophisticated approach to personalized medicine, it faces inherent limitations. Causal models derived from observational data are notoriously difficult to validate without prospective trials. Furthermore, observational bias remains a concern; if a key confounder—such as a patient's lifestyle or a specific unrecorded symptom—is missing from the database, the model's recommendations could be biased.

Future work will focus on the generalizability of the tool, ensuring that the strategies developed using the Stanford dataset can be effectively applied across different hospital systems and diverse patient populations. By bridging the gap between causal inference and clinical decision-making, this research moves us closer to a future where cancer treatment is as dynamic and personalized as the disease itself.

Summary updated

BibTeX
@inproceedings{zeng2020developing,
  author = {Jiaming Zeng},
  title = {{Developing a Machine Learning Tool for Dynamic Cancer Treatment Strategies}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i10.7143},
}
Figure