TLDR: This research introduces two deep learning frameworks to improve the diagnosis and data completeness for ocular-induced abnormal head posture (AHP). AHP-CADNet, a multi-level attention fusion model, integrates ocular landmarks, head pose features, and clinical attributes for automated diagnosis, achieving high accuracy (96.9%–99.0%) and low prediction errors. The second framework uses curriculum learning and a Transformer-based model (PubMedBERT) to impute missing clinical data from structured variables and unstructured notes, maintaining high accuracy (93.46%–99.78%) and showing significant gains from clinical dependency modeling. These frameworks offer a robust approach to objective AHP diagnosis and reliable data recovery in clinical settings.
Ocular-induced abnormal head posture (AHP) is a condition where individuals adopt unusual head positions to compensate for eye misalignment, such as strabismus. This compensatory mechanism helps them reduce double vision and maintain clear binocular vision. Early and accurate diagnosis of AHP is crucial to prevent long-term complications like facial asymmetry, neck pain, and spinal issues. However, current clinical assessments are often subjective and complicated by incomplete patient medical records.
A recent study, titled Ocular-Induced Abnormal Head Posture: Diagnosis and Missing Data Imputation, addresses these challenges by introducing two innovative deep learning frameworks. These frameworks aim to automate diagnosis and effectively manage missing data in clinical records, enhancing the reliability and objectivity of AHP assessment.
Automated Diagnosis with AHP-CADNet
The first framework, AHP-CADNet (Abnormal Head Posture – Clinical Attention Diagnosis Network), is designed for automated diagnosis of ocular-induced AHP. It uses a sophisticated multi-level attention fusion approach that integrates various types of patient data: ocular landmarks (features from the eyes), head pose features, and structured clinical attributes (like reported symptoms and diagnostic labels). By combining these diverse data sources, AHP-CADNet can generate interpretable predictions for both the underlying ocular misalignment and the resulting head posture.
The model performs several tasks, including classifying the type of AHP, identifying the affected eye, diagnosing the specific ocular condition, and predicting continuous variables like prism diopter (PD) values (a measure of eye deviation) and the angular degree of head deviation. The multi-level attention mechanism allows the model to understand relationships within each data type (e.g., how different eye landmarks relate) and across different data types (e.g., how clinical symptoms relate to eye features or head posture).
Evaluated on the PoseGaze-AHP dataset, AHP-CADNet demonstrated robust diagnostic performance. It achieved high accuracy, ranging from 96.9% to 99.0% across various classification tasks, and low prediction errors for continuous variables, with Mean Absolute Error (MAE) between 0.103 and 0.199, and R² values exceeding 0.93. This indicates that the framework is highly effective in accurately identifying and characterizing ocular-induced AHP.
Addressing Missing Data with Curriculum Learning
The second framework tackles the common problem of incomplete medical records in electronic health record (EHR) systems. It’s a curriculum learning-based imputation framework designed to fill in missing data by progressively leveraging both structured variables and unstructured clinical notes. This is particularly important because critical details are often found in narrative clinical notes, which traditional imputation methods might overlook.
The framework uses a progressive masking strategy, informed by clinical knowledge, to gradually increase the difficulty of imputation tasks. It also employs a pre-trained biomedical language model, such as PubMedBERT, to understand the context and terminology within clinical notes. By modeling clinical dependencies—meaning certain clinical features predict others—the framework ensures that the imputed data is clinically relevant and accurate.
This imputation framework maintained high accuracy across all clinical variables, achieving between 93.46% and 99.78% accuracy when using PubMedBERT. Furthermore, incorporating clinical dependency modeling led to statistically significant improvements (p < 0.001), confirming its effectiveness in recovering missing data under realistic clinical conditions. The framework showed robust performance even for targets with high rates of missingness, such as AHP degree and PD, demonstrating its ability to generalize from partially observed to fully imputed scenarios.
Also Read:
- Unveiling PoseGaze-AHP: A New 3D Dataset for AI-Powered Eye and Head Posture Diagnosis
- Advancing Medical AI with MedCLM: A Curriculum for Visual Reasoning and Localization
Impact and Future Directions
The findings from this study confirm the effectiveness of both proposed frameworks for automated diagnosis and recovery from missing data in clinical settings. By integrating multimodal data and employing advanced deep learning techniques, these approaches hold significant promise for enhancing the objectivity and reliability of ocular-induced AHP diagnosis.
While the results are promising, the authors acknowledge limitations, including the reliance on a simulated dataset and the need for validation with larger, more diverse real-world EHRs. Future research will focus on validating these frameworks with broader datasets, extending the diagnostic capabilities to other AHP causes, and integrating them into commercial EHR systems for prospective clinical trials.
