Introduction
Developmental delays affect approximately 6% to 9% of children globally. Early identification is paramount, as the "golden period" for intervention occurs before the age of three and a half. During this window, targeted treatments can lead to significant improvements or even full recovery. However, traditional screening processes are often hampered by subjectivity in bone-age assessment, a scarcity of professional resources, and the inherent complexity of evaluating motor, language, and cognitive growth.
This paper presents a hybrid diagnostic system designed to automate and standardize developmental screening. By combining deep learning for objective bone-age assessment with Case-Based Reasoning (CBR) for holistic developmental analysis, the system provides a robust decision-support tool for clinicians.
Methodology: A Hybrid Diagnostic Framework
The proposed system utilizes a two-pronged approach to evaluate a child's developmental status.
Bone Age Assessment via CNN-Transformers
To provide an objective measure of physical maturity, the system analyzes hand X-rays using a hybrid deep learning architecture.
- Feature Extraction (CNN): A pre-trained DenseNet architecture captures local morphological details of the bones, while Xception is employed for fine-tuning.
- Global Modeling (Transformer): To complement the CNN's local focus, a Transformer block utilizing a Fast-Attention Mechanism captures long-distance dependencies across the entire skeletal structure. This ensures a comprehensive prediction of bone age rather than relying on isolated features.
Case-Based Reasoning (CBR)
The system mimics human clinical reasoning through CBR, which solves new problems by referencing successful historical diagnoses. This process follows the "4R" cycle:
- Retrieve: The system searches a database for cases with similar physiological and developmental profiles.
- Reuse/Adapt: Solutions from these historical cases are applied to the current patient.
- Revise: A senior clinician (with over 20 years of experience) reviews and, if necessary, corrects the AI-generated suggestion.
- Retain: The verified case is added to the database, allowing the system to improve its knowledge base over time.
A weighted similarity algorithm compares cases across six key categories: physiological indicators, language, social interaction, gross motor skills, fine motor skills, and cognitive development. Notably, "Actual Age" is given the highest similarity weight (20), while other developmental milestones are weighted equally (8).
Datasets and Preprocessing
The study utilized a primary dataset of 1,342 hand X-ray images obtained from the pediatric department of Huaibei People's Hospital in 2023.
The case database initially consisted of 1,233 records. For system validation, a subset of 100 cases from 2023 was used, while a separate verification set of 50 cases from 2022 was used to test diagnostic performance. Image preprocessing included resizing to 224x224 pixels, normalization, and data augmentation via rotation, flipping, and Generative Adversarial Networks (GANs) to enhance model robustness.
Key Insights: The "I Don't Know" Metric
A unique feature of the screening scale is the inclusion of an "I Don't Know" option for caregivers. The researchers use this as a reliability filter; if more than 10% of responses fall into this category, the assessment is flagged as unreliable. This may indicate a lack of observation by the caregiver or potential psychological distress, prompting further clinical investigation.
Results and Performance
The system's performance was validated against established clinical outcomes:
- Average Similarity: 0.9217 (92%), demonstrating the system's ability to retrieve highly relevant precedent cases.
- Reasoning Accuracy: 0.9129 (91%), showing that the system's diagnostic suggestions align closely with professional expert diagnoses.
- Efficiency: The automated generation of screening charts and results significantly reduces the manual workload for medical professionals.
Discussion and Conclusion
While the system shows high accuracy, it is currently limited by its single-center data source. Future work will focus on expanding the dataset to include diverse demographics and optimizing database search speeds as the number of cases grows. Additionally, while the "Revision" stage ensures safety, reducing dependency on senior experts will be crucial for broader scalability.
By bridging the gap between physiological growth (bone age) and functional development (behavioral screening), this hybrid model offers a promising path toward more efficient and objective early intervention for children with developmental delays.
Summary updated