Source-linked AI summary
Feature importance analysis for patient management decisions
Michal Valko, Milos Hauskrecht
TL;DR
The paper asks which clinical-data characteristics most influence physicians’ laboratory and medication decisions. It analyzes temporal features from electronic health records of post-surgical cardiac patients using AUC-based feature significance and limited multivariate models. The results indicate that simple patterns and a small number of clinical variables often predict these decisions well.
Problem
The study addresses which temporal features of complex multivariate clinical time series most influence physicians’ lab-order and medication decisions.
Method
The authors analyze feature importance in records from 4486 post-surgical cardiac patients using AUC scores, category summaries, and limited linear-SVM models based on top-ranked features.
Results
Lab and medication order decisions are often driven by simple patterns involving recent values or time since events, and limited analyses suggest dependence on only a few clinical variables.
Takeaways & Limitations
Simple temporal patterns and their refinements may capture many prevalent lab-order and medication-order patterns.
Takeaways & Limitations
The multivariate experiment was limited, and greedy feature selection among over 9000 candidates may not have been optimal.
Abstract
from arXiv · showhide
The objective of this paper is to understand what characteristics and features of clinical data influence physician's decision about ordering laboratory tests or prescribing medications the most. We conduct our analysis on data and decisions extracted from electronic health records of 4486 post-surgical cardiac patients. The summary statistics for 335 different lab order decisions and 407 medication decisions are reported. We show that in many cases, physician's lab-order and medication decisions can be well predicted from a small subset of all features.
Introduction
This paper examines which temporal characteristics of patient data most influence physicians’ lab-order and medication decisions. Using 4486 post-surgical cardiac patient records, it finds that relatively simple temporal characterizations often predict many such decisions well.
- Introduction: The study addresses which temporal features physicians use when deciding whether to order laboratory tests or medications.It considers recency, values, trends, time constraints, and relationships among clinical variables.
- Introduction: The analysis focuses on feature importance for both lab-order and medication-order decisions.
- Introduction: 4486 post-surgical cardiac patient records were analyzed to study these decisions.
- Introduction: Relatively simple temporal characterizations often predict many lab-order and medication decisions well.The paper also identifies which simple characteristics provide the most valuable predictive information.
PCP Dataset
The study uses de-identified records from 4486 post-surgical cardiac patients, converting time-stamped clinical data into patient-state examples and linking each state to decisions made during the following 24 hours. These states are represented with temporal features from laboratory, medication, demographic, procedure, and heart-support-device data.
- PCP Dataset: The PCP database contains de-identified records for 4486 post-surgical cardiac patients treated at a UPMC teaching hospital.Records include demographics, notes, laboratory and test results, medications, procedures, and other clinical data.
- PCP Dataset: Each patient record was converted into a sequence of patient-state examples representing information available at 8:00am every 24 hours.Only information available before each segmentation point was included in the state representation.
- PCP Dataset: Each patient state was linked to lab-order and medication decisions made during the next 24 hours.The decisions record whether each lab was ordered or each medication was given.
- PCP Dataset: The representation uses features summarizing measurements, trends, medication status, demographics, procedures, and heart-support-device use.Laboratory features include recent values, temporal changes, and time since measurement; other categories encode medication, demographic, procedure, and device status.
- PCP Dataset: The dataset contains 9,223 features describing 30,828 patient states and evaluates 742 lab and medication order decisions.
Methods
The paper evaluates how strongly individual feature categories predict lab and medication order decisions, using AUC-based feature significance and summary statistics across many decisions. A limited multivariate analysis then tests whether small sets of top-ranked features improve prediction over single features.
- Methods: The objective is to evaluate feature significance for predicting lab-order or medication-order decisions.
- Methods: 9,223 features were grouped into five clinical categories and 40 temporal feature categories for analysis across multiple decisions.The temporal categories represent shared characteristics such as time since the last laboratory measurement.
- Methods: AUC measured each feature’s predictive strength, while category importance counted how often a category contained the best-AUC feature.
- Methods: The limited multivariate analysis trained linear SVM models using the top 1, top 3, or top 30 AUC-ranked features for each decision.Performance was assessed on remaining patient cases after training on a subset of 2900 patients.
Results
Lab-order decisions were driven mainly by laboratory and procedure features, while medication commissions were dominated by procedure features. Temporal patterns also showed that simple recent values, elapsed times, and medication co-occurrence often predicted decisions effectively.
- Lab-order decisions: Lab and procedure features were the most influential predictors of lab-order decisions.Past lab results and procedure context were especially informative for predicting the next lab order.
- Lab-order decisions: Time since the last lab order or result, along with last value, nadir, and horizon values, were the strongest temporal categories for ordering the same lab.Absolute lab values were generally more important than trends, with trends typically refining value-based patterns.
- Lab-order decisions: For ordering a different lab, the last value of another lab was the most significant predictor.Abnormal results and joint ordering of laboratory panels were identified as explanations for this dependency.
- Medication decisions: Procedure features dominated medication-commission prediction, with some medications following particular procedures.For Papaverine, time since the last coronary artery bypass predicted commissions with AUC=95%.
- Medication decisions: Medication features from other drugs were influential because complementary medications may be administered together, while some combinations are avoided because of possible drug interactions.The presence of one medication could therefore help predict the presence or absence of another.
- Multivariate prediction: A linear SVM experiment evaluated top 1, top 3, and top 30 selected features for predicting highly predictive labs and medications on an independent test set.The authors note that greedy feature selection may not be optimal with over 9000 feature candidates, although a single feature was often a good predictor.
Conclusions
The analyses found that lab and medication decisions are often driven by simple, recent temporal patterns involving values and elapsed time since clinical events. Limited multivariate analyses further suggest that only a few clinical variables and their features may underlie many decisions.
- Recent values and elapsed time since procedures or prior orders often form simple predictive patterns for lab and medication decisions.
- Limited multivariate analyses suggest that lab and medication decisions are likely based on only a few clinical variables and their characteristics.
- The planned next steps are to analyze feature dependencies and develop more advanced feature selection for multivariate predictive models.