Introduction
The early detection of pathologic cognitive decline, such as dementia, is critical for improving the quality of life for the elderly. While digital cognitive assessments—analyzing motor patterns and touchscreen interactions—offer a cost-effective alternative to traditional clinical evaluations, they are often hindered by the cross-task gap. This phenomenon occurs when a model trained on one device (e.g., a 10-inch tablet) or task fails to generalize to another (e.g., a 21-inch monitor) due to shifts in data distribution. Given the difficulty of collecting large-scale labeled medical data, robust transfer learning methods are essential to bridge these gaps.
The Fine-Grained Adaptation Random Forest (FAT)
To address these challenges, researchers developed the Fine-Grained Adaptation Random Forest (FAT). Unlike traditional transfer learning methods that treat models as monolithic entities, FAT operates at the level of individual decision trees. The framework consists of two core components:
- Information Gain based Model Evaluation (IGME): This component "grades" each tree in the source forest by calculating its relevance to the target domain. It identifies which decision paths remain valid and which have been compromised by the domain shift.
- Domain Adaptation Tree Growing (DATG): Based on the IGME scores, the model applies specific remodeling strategies to the trees to align them with the target data distribution.
Adaptation Strategies
The DATG stage employs three primary strategies to align the source model with the target domain:
- Modify The Subtree (MTS): If a specific branch performs poorly due to structural misalignments, it is pruned and retrained using target data.
- Split Leaf Nodes (SLN): To capture finer nuances in more complex target data, the model grows additional leaves on existing branches.
- Update Feature Threshold (UFT): For shifts in scale—such as longer swipe distances on larger screens—the model adjusts the numerical thresholds at decision nodes.
Experimental Results
The FAT method was evaluated in real-world clinical environments using two distinct hardware setups: a Huawei M5 tablet (10.1 inches) and a NanoPi M4 board (21.5 inches). Tasks included digital "Box and Block" tests in both single-task and dual-task (cognitive-motor) configurations.
FAT demonstrated significant superiority, achieving the highest accuracy in 55 out of 56 transfer task combinations. It reached a peak diagnostic accuracy of 92.8% and maintained high performance even when using only 30% of the target data for fine-tuning. This efficiency is crucial for medical applications where labeled data is scarce and expensive to obtain.
Conclusion
The Fine-Grained Adaptation Random Forest provides a robust solution for cross-task cognitive assessment. By surgically adapting decision trees rather than retraining from scratch, FAT preserves valuable knowledge while effectively bridging the distribution gap. Its high accuracy and interpretability make it a promising tool for clinical deployment. Future research will focus on unsupervised adaptation techniques to further minimize the need for labeled target data.
Summary updated