README – CARE-MIRV-Net Supplementary Material
Overview

This supplementary package contains the implementation files associated with the manuscript entitled:

"CARE-MIRV-Net: A Healthcare-Oriented Residential Layout Assessment Framework Using Deep Learning, Meta-Ensemble Learning, and Explainable Artificial Intelligence"

The package provides the source code and supporting files required to reproduce the experimental workflow presented in the manuscript.

Dataset Information

Dataset Name: Floor Plan Images and Their Details

Source: Kaggle

Dataset Link:
https://www.kaggle.com/datasets/adilmohammed/floor-plan-images-and-their-details

The dataset consists of residential floor-plan images and associated metadata, including square footage, number of bedrooms, bathrooms, and garages. These attributes are used as target variables for model development and evaluation.

Contents of Supplementary File 1
1. CARE_MIRV_Net.py

Main implementation of the proposed CARE-MIRV-Net framework. The script includes:

Data loading and preprocessing
Training of MobileNetV2
Training of InceptionV3
Training of ResNet101
Training of VGG16
Meta-ensemble model construction
Model evaluation using MAE and R² metrics
SHAP-based explainability analysis
Residential categorization and decision-support procedures
2. requirements.txt

This file contains the software dependencies and library versions required to execute the CARE-MIRV-Net framework.

The experimental environment was developed using:

Python 3.12.12
TensorFlow 2.19.0
Keras 3.13.2
NumPy 2.4.6
Pandas 2.3.3
Scikit-learn 1.6.1
Matplotlib 3.9
SHAP 0.51.0

Installing the listed dependencies helps ensure consistency and reproducibility of the reported experiments.

Framework Description

The CARE-MIRV-Net framework employs four deep convolutional neural networks:

MobileNetV2
InceptionV3
ResNet101
VGG16

Each model is trained independently to predict key residential floor-plan attributes:

Square footage
Number of bedrooms
Number of bathrooms
Number of garages

The predictions generated by the individual models are subsequently combined using a stacking-based meta-ensemble strategy to improve predictive robustness and generalization performance.

An interpretation layer is further applied to transform the predicted residential attributes into application-oriented residential assessment categories.

To improve transparency and interpretability, SHapley Additive exPlanations (SHAP) are employed to quantify the contribution of the base models to the final predictions.

Performance Evaluation

Model performance is evaluated using:

Mean Absolute Error (MAE)
Coefficient of Determination (R²)

These metrics provide complementary assessments of prediction accuracy and explanatory capability across all target variables.

Hardware and Software Environment

The experiments were conducted in a cloud-computing environment using:

Intel® Xeon® CPU @ 2.20 GHz
NVIDIA Tesla T4 × 2 GPUs
Approximately 31 GB RAM
Ubuntu 22.04 LTS operating system
Reproducibility

The files included in this supplementary package are provided to support the reproducibility of the experimental procedures and results reported in the manuscript. Users may adapt the implementation for further research and validation purposes using the publicly available dataset referenced above.