Imputing ACOG Scores with AI Model

Table of Contents

1. Introduction
2. Background on ACOG scores
3. Challenges in incorporating real-world patient data
4. The need for an AI model to impute ACOG scores
5. The research abstract and its significance
6. Methods used to build the machine learning model
7. Features used in the model
8. Results and evaluation of the data-driven model
9. Results and evaluation of the selected feature model
10. Interpretation of the results and implications for clinical trials
11. Conclusion and future research

An AI Model to Impute ACOG Scores in Real-World Patient Data

Introduction

In the field of oncology, Accurate Collection of Outcome and Generalizability (ACOG) scores play a crucial role in assessing cancer prognosis and determining patient eligibility for clinical trials. However, in real-world clinical settings, these scores are often missing from the structured and unstructured data of patients who are not part of clinical trials. This poses a significant challenge in using real-world data to establish external control arms. In this article, we will explore an innovative AI model developed to impute ACOG scores using machine learning techniques.

Background on ACOG Scores

ACOG scores are powerful prognostic indicators of cancer outcomes. They are typically assessed on a Scale from zero to one, with lower scores indicating more favorable prognoses. These scores are frequently required for clinical trial enrollment, as they help identify patients with specific prognostic characteristics. However, obtaining ACOG scores in real-world settings is often challenging, as patients treated outside of clinical trials may not have these scores documented in their medical records.

Challenges in Incorporating Real-World Patient Data

The absence of ACOG scores in real-world patient data presents several challenges. Firstly, it limits the ability to include real-world patients in external control arms, which are essential for comparative effectiveness research. Secondly, it hampers the analysis of real-world data to understand the outcomes and effectiveness of different treatments. To address these challenges, an AI model was developed to impute ACOG scores Based on available information from different points in the patient Journey.

The Need for an AI Model to Impute ACOG Scores

To overcome the limitations of missing ACOG scores in real-world patient data, a machine learning model was developed. This model aims to predict ACOG scores using various patient-related features, including previous ecoegg information, future encounters, medications, latest lab values, and more. By imputing ACOG scores, this AI model enables the incorporation of real-world data into external control arms, thereby improving the accuracy and validity of clinical trial studies.

The Research Abstract and its Significance

The AI model for imputing ACOG scores was developed based on a curated NSCLC (non-small cell lung cancer) cohort of 30,1425 patients. The XGBoost classifier was utilized as the highest-performing machine learning technique, with a logistic regression baseline. The abstract presented at the ASCO 2020 virtual scientific program highlighted the success of the model in predicting ACOG scores and discussed its implications for clinical trial studies and external control arms.

Methods Used to Build the Machine Learning Model

The machine learning model employed a data-driven approach, utilizing over 22,000 features extracted from the curated NSCLC cohort data. These features included previous ecoegg information, future encounters, medications, latest lab values, age, stage, number of comorbidities, and transformations of key labs. Additionally, a selected feature approach focused on using only past data and a restricted set of features to Create a more interpretable model.

Features Used in the Model

The machine learning model incorporated a wide range of features to predict ACOG scores. By considering past data and a comprehensive set of patient-related information, the model aimed to capture the complex relationship between various factors and ACOG scores. Features such as age, stage, comorbidities, and lab values played a crucial role in the prediction process.

Results and Evaluation of the Data-Driven Model

The data-driven model demonstrated high accuracy in imputing ACOG scores. The area under the curve (AUC) was found to be 0.81, indicating a reliable prediction of ACOG scores. The precision-recall curve (PRC) showed an average precision of 0.87 for good ACOG scores and an average precision of 0.72 for poor ACOG scores. These results suggest that the model can effectively identify patients with favorable ACOG scores, making it a valuable tool for clinical trial enrollment and external control arms.

Results and Evaluation of the Selected Feature Model

The selected feature model, which focused on using a restricted set of features, achieved similar results to the data-driven model. The AUC was slightly lower at 0.77, indicating a slightly reduced accuracy compared to the data-driven approach. However, the average precision for good ACOG scores remained high at 0.83, and for poor ACOG scores, it was 0.674. These findings suggest that a more focused and interpretable model can achieve comparable performance to a more comprehensive model.

Interpretation of the Results and Implications for Clinical Trials

The high accuracy and performance of both machine learning models indicate their potential for use in clinical trials. By imputing ACOG scores accurately, these models enable the creation of external control arms comprising real-world patients. This enhances the robustness and generalizability of clinical trial studies, allowing for better assessment of treatment outcomes and effectiveness. However, further research is warranted to compare the outcomes of patients predicted by the machine learning models with those assigned ACOG scores by medical professionals.

Conclusion and Future Research

The development of an AI model to impute ACOG scores in real-world patient data is a significant step towards leveraging machine learning in the field of oncology. The results demonstrate the accuracy and potential of these models in predicting ACOG scores, facilitating the inclusion of real-world patients in clinical trials. Further research should focus on evaluating the outcomes and impact of using machine learning-predicted ACOG scores in various clinical scenarios. This research opens up new avenues for optimizing clinical trial designs and enhancing patient-centered research in the field of oncology.