AI-Fairness in Elder Care: Decoding Activity Recognition Across Physical Disabilities

TL;DR

Recognizing the daily activities of older adults using a single wearable is notoriously difficult due to "functional diversity"—the high variance in how someone walks with a walker versus a cane. This paper proposes a dual-path solution: a hardware-efficient Sparse-Deconvolution MM-RNN architecture for multi-label recognition and a rigorous AI-Fairness pipeline to ensure the model doesn't discriminate against the oldest or most disabled users.

Problem & Motivation: The Hidden Bias in Wearables

Standard Human Activity Recognition (HAR) assumes a "gold standard" for movement. However, for a 90-year-old using a walker, "walking" looks mechanically similar to "standing" or "shuffling" to a traditional AI. This leads to algorithmic unfairness: the person who needs the most monitoring (the disabled) receives the least accurate detection.

The author identifies two core gaps:

1. Sensor Fatigue: Older adults refuse to wear multiple sensors; we must extract Upper Extreme (UE) and Lower Extreme (LE) data from a single wristband.
2. Diversity Bias: Age and functional ability act as "protected attributes" that skew AI performance, potentially violating legal standards like the US Disability Discrimination Act.

Methodology: Physics-Inspired Signal Separation

The paper treats the human body as a Linear Time-Invariant (LTI) system. The core technical insight is that Lower Extreme (LE) movements (like walking) have a "sparse effect" on Upper Extreme (UE) nodes (the wrist).

1. Sparse Deconvolution

By assuming the wrist signal is a convolution of a hand gesture and a postural response, the author uses a sparse-deconvolution method to disaggregate the two. This allows a single sensor to "see" what the legs are doing while simultaneously tracking hand gestures.

2. MM-RNN Architecture

The separated signals are processed through a Multi-Input Multi-Task Recurrent Neural Network (MM-RNN).

• Input: Two time-series streams (Gestural + Postural).
• Middle: Three stacked LSTM layers for temporal dependencies.
• Output: Dual classification heads for concurrent activity labeling.

Figure 1: The architecture transition from signal deconvolution to deep multi-task learning.

Mitigating Bias: Fairness as a First-Class Citizen

The most significant contribution is the systematic evaluation of Bias Mitigation. The author tested three intervention points:

• Pre-processing: Reweighing samples to balance groups.
• In-processing: Adversarial debiasing to prevent the model from "learning" age-based shortcuts.
• Post-processing: Adjusting the decision boundary (Reject Option Classification).

Figure 2: Evaluating different algorithms (Reweighing vs. Prejudice Remover) on fairness metrics like Statistical Parity and Equal Opportunity.

Experiments & Results

The study involved 22 participants (mean age 85.5) in a retirement community.

• Raw Performance: The MM-RNN achieved 97.5% accuracy, beating Feature Weighted Naive Bayes (FWNB) and SVMs.
• Fairness Gains: For the protected attribute "Age," the Reject Option algorithm achieved a "4 out of 4" fairness score (perfect across all metrics) with a negligible 0.5% accuracy loss.
• Challenge of walking aids: Accuracy for "Normal Walking" was near 100%, but dropped significantly for "Walking with Walker" (approx. 80.7%), highlighting why specialized bias mitigation is necessary.

Critical Insight & Conclusion

This work proves that Accuracy is not a proxy for Utility in healthcare. A model that is 95% accurate across the general population but 60% accurate for disabled users is fundamentally broken for clinical deployment.

Takeaway: By combining signal deconvolution (to maximize information from sparse sensors) and post-processing fairness adjustments, we can build HAR systems that are both high-performing and ethically sound. Future work focusing on Explainable AI (XAI) will be the next step in helping clinicians understand why a certain gait was flagged, further bridging the gap between data science and geriatric care.