Accurate and Efficient Real-World Fall Detection Using Time Series Techniques

Abstract

Falls pose a significant health risk, particularly for older people and those with specific medical conditions. Therefore, timely fall detection is crucial for preventing fall-related complications. Existing fall detection methods often have high false alarm or false negative rates, and many rely on handcrafted features. Additionally, most approaches are evaluated using simulated falls, leading to performance degradation in real-world scenarios. This paper explores a new fall detection approach leveraging real-world fall data and state-of-the-art time series techniques. The proposed method eliminates the need for manual feature engineering and has efficient runtime. Our approach achieves high accuracy, with false alarms and false negatives each as few as one in three days on FARSEEING, a large dataset of real-world falls (mean F$_1$ score: 90.7%). We also outperform existing methods on simulated falls datasets, FallAllD and SisFall. Furthermore, we investigate the performance of models trained on simulated data and tested on real-world data. This research presents a real-time fall detection framework with potential for real-world implementation.

Appendices

Appendix A Usability and Explainability

1.1 Appendix A.1 Estimated False Alarm and Miss Rates

Since we use an overlapping sliding window with a step size of 1 s, each classifier would process a total of $1 \times 60 (\text {s}) \times 60 (\text {min}) = 3600$ samples per hour in practice. Taking the Hydra classifier as an example, we obtain a rough estimate of false alarm and miss rates per hour (Table 5) in the following manner.

Let P be the total number of falls, and N be the total number of ADLs per hour, such that $P + N = 3600$. From the Fall: ADL ratio in Table 1, we know that

$$\begin{aligned} P = 0.14N \end{aligned}$$

(1)

So that

$$\begin{aligned} 0.14N + N = 3600 \end{aligned}$$

(2)

Hence, we can estimate the number of ADLs per hour as $N=\frac{3600}{1.14} \approx 3158$, and the number of falls $P=3600-3158=442$ per hour. Therefore, we estimate the number of misses per hour, FN as

$$\begin{aligned} FN = P \times (1 - Recall) \end{aligned}$$

(3)

$$\begin{aligned} FN = 442 \times (1-0.8913) \end{aligned}$$

(4)

$$\begin{aligned} FN \approx 48 \end{aligned}$$

(5)

Similarly, false alarms per hour, FP:

$$\begin{aligned} FP = N \times (1 - Specificity) \end{aligned}$$

(6)

$$\begin{aligned} FP = 3158 \times (1-0.986) \end{aligned}$$

(7)

$$\begin{aligned} FP \approx 44 \end{aligned}$$

(8)

Hence, miss rate $=\frac{48}{3600} \approx 0.013$, and false alarm rate $= \frac{44}{3600} \approx 0.012$ per hour.

Table 5. False Alarm Rate and Miss Rate Per Hour

1.2 Appendix A.2 Model Explanation

We use tscaptum [7, 44] to identify time intervals in motion data that influence classifier decisions on FARSEEING. tscaptum groups adjacent time points into segments of size c to enhance robustness and reduce runtime, with important intervals identified by iteratively masking segments. SHAP (SHapley Additive exPlanations) scores [27] are then obtained based on their impact on predictions.

We perform a subject-wise train-test split, using 30% (13 subjects) for testing and 70% (28 subjects) for training (see Fig. 7). Temporal SHAP scores are obtained on the test set with $c=100$, representing a 1-s interval (sampled at 100 Hz). This produces equally distributed attribution scores within each chunk, matching the shape of the input sample.

We present attribution profiles for representative true positives, false positives, true negatives, and false negatives across all classifiers. Although the impact phase occurs within $t=[1,2)$, attribution scores for Hydra (Fig. 8) and MultiRocketHydra (Fig. 10) suggest uniform importance across all phases. However, Rocket (Fig. 9), Catch22 (Fig. 11), and QUANT (Fig. 12) show high scores between the end of the falling phase and the start of the impact phase.

Fig. 7.

Confusion matrices for time series classifiers on the FARSEEING test set.

Fig. 8.

Hydra + SHAP temporal attribution profiles for some samples.

Fig. 9.

Rocket + SHAP temporal attribution profiles for some samples.

Fig. 10.

MultiRocketHydra + SHAP temporal attribution profiles for some samples.

Fig. 11.

Catch22 + SHAP temporal attribution profiles for some samples.

Fig. 12.

QUANT + SHAP temporal attribution profiles for some samples.

1.3 Appendix A.3 Model Runtime

In our experiments, we compute runtime in milliseconds (ms) as the time it takes each model to perform inference on one sample. As shown in Fig. 13, the tabular models run extremely fast. However, the time series models also run fast enough to be deployed in real-time, with MultiRocketHydra having the slowest inference speed of 144 ms per sample in one instance.

Fig. 13.

Inference time per sample in milliseconds for all models.

Appendix B Preprocessing Choices

1.1 Appendix B.1 Univariate vs Multivariate

We performed a single train/test split experiment on FARSEEING to compare the performance of univariate acceleration magnitude with multivariate x, y, z acceleration signals. As shown in Fig. 14, using univariate magnitude signals with the time series classifiers showed significantly better scores in all metrics.

Fig. 14.

Univariate acceleration magnitude vs. multivariate x, y, z acceleration on FARSEEING. Each box represents the performance of all time series classifiers.

1.2 Appendix B.2 Window Size and Post-fall Phase

We trained the time series models on the FARSEEING dataset with total sample window sizes ranging from 3 s (1-s post-fall phase) to 27 s (25-s post-fall phase). As shown in Fig. 15, longer post-fall phases improve AUC but don’t necessarily improve recall, specificity, or F$_1$ scores. In fact, they may degrade performance and increase runtime. Experts recommend that a 5-s post-fall phase is sufficient to capture the necessary context.

Fig. 15.

Effect of post-fall window size on model performance on the FARSEEING dataset.

Appendix C Extended Cross-Validation Results

Here, we give more detailed results, including cross-validation split sizes and the performance of each classifier on individual splits. We report the inference time per sample (T) in milliseconds, AUC, precision, recall, specificity, and F$_1$ score.

Table 6. Cross-validation splits

Table 7. Performance of Tabular Models on the FARSEEING Dataset

Table 8. Performance of Tabular Models on the FallAllD Dataset

Table 9. Performance of Tabular Models on the SisFall Dataset

Table 10. Performance of Time Series Models on the FARSEEING Dataset

Table 11. Performance of Time Series Models on the FallAllD Dataset

Table 12. Performance of Time Series Models on the SisFall Dataset

Table 13. Results of Cross-Dataset Evaluation