Dose-report-driven body weight estimation in diagnostic CT using a practical machine learning framework

Area of Science:

• Radiology and Medical Imaging
• Health Informatics
• Machine Learning in Healthcare

Background:

• Patient body weight is crucial for accurate contrast media and radiopharmaceutical dosing, radiation dose management, and optimizing examination workflows in radiology.
• However, patient body weight is not consistently recorded or readily available in routine clinical settings, posing challenges for these critical applications.

Purpose of the Study:

• To assess the feasibility of estimating adult body weight by analyzing diagnostic computed tomography (CT) dose report metrics across different CT systems.
• To develop and validate machine learning models for predicting patient weight using readily available CT data.

Main Methods:

• A retrospective study analyzed data from 2496 adult patients undergoing diagnostic CT on three different scanners, using measured body weight as the reference standard.
• LightGBM regression models were trained and evaluated using dose report-derived metrics (e.g., water-equivalent diameter, sex), body region, and CT system information.
• Model performance was assessed using mean absolute error (MAE) and the percentage of estimates within ±10% of the reference weight, including cross-scanner validation.

Main Results:

• The baseline model, using patient size and exposure metrics, achieved an MAE of 3.07 kg, with 86.0% of estimates within ±10% of the reference weight.
• The extended model, incorporating body region and CT system, showed a slight improvement with an MAE of 2.83 kg and 88.8% of estimates within ±10%.
• Scanner-wise cross-validation yielded MAE values ranging from 3.16 kg to 4.37 kg, indicating reasonable generalizability across different systems.

Conclusions:

• Diagnostic CT dose report information can be utilized to estimate adult body weight with practical accuracy.
• This approach offers a viable solution for obtaining essential weight data in clinical radiology, supporting dose management and workflow optimization.
• The findings suggest that machine learning models leveraging CT dose metrics can enhance data availability for critical patient care decisions.