*2.4. Classification Strategies*

The objective of FDSs is to identify whether a fall happened or not, hence a binary decision. FDSs previously reported in the literature typically follow a binary classification approach, following the intuition that the event of interest is whether the participant has fallen or not. A notable exception to this trend [25] extends this common approach by aiming to differentiate amongs<sup>t</sup> various causes of falls. The study differentiates three causes of falls which are trips, slips and others. Another study [23] used a different approach where the type of fall is identified (amongst the types forward, backward, lateral) as well as various ADLs. In a different context, which is Fall Prevention System [38], the goal is to detect if a fall will definitively happen in order to deploy a protection mechanism such as airbags. In such systems, it is not the fall that needs to be detected but what we could call the pre-fall, meaning what happens before the actual fall. More recent approaches combine these two ideas, for example [32] used a multi-phase model. They differentiate phases of a fall and then classify them into three classes: free fall, impact and rest phases. We further investigate this, using several ML algorithms as detailed in Section 3.4.

### *2.5. Strengths and Weaknesses*

Wearable technologies have several advantages. They are relatively inexpensive and can operate anywhere all of it with minimal intrusion compared to other approaches, such as environmental monitoring [33,34]. In addition, their somewhat limited computational power can be easily overcome with the use of their telecommunication capabilities, which allow the transfer of data for processing outside the device. Wearables can also

identify the wearer and ge<sup>t</sup> precise measurements. However, they may create discomfort due to their size and intrusiveness. The main disadvantage is their human dependency. These sensors must have enough battery and be worn to work properly. Furthermore, the elderly may have a cognitive impairment and thus, may forget to wear the sensor.

### **3. Materials and Methods**

Our FDS is based on a common pipeline (Figure 2) which has been seen in the literature [27]. This pipeline is a common practice when working with ML algorithms. We first acquire raw data using various sensors and convert them into discrete values. We then preprocess the raw data to remove measuring errors which can badly affect the performance. Afterwards, we construct and extract meaningful information in a vector. Finally, we train and evaluate our ML algorithm to distinguish falls from ADLs.

**Figure 2.** General architecture of Fall Detection Systems.

The steps presented above for our FDS pipeline are common to most of our research questions. However, in order to address research question 3, we have a few changes that will be highlighted. Thus, this section is organised as follows: Sections 3.1 to 3.5 details each step of the pipeline which are common to all research questions. These five subsections answers entirely the first two research questions. However, the third research question requires additional data preparation which is described in Section 3.6.
