**6.3 Algorithms, artificial intelligence and machine learning**

Algorithms, artificial intelligence (AI) and machine learning (ML) have been slow to be applied to telehealth. Most telehealth systems continue to use simple thresholds on vital signs data to determine patients at risk of deterioration, which has been shown to have poor performance, resulting in such frequent false alerts that they are often disabled by users and alerts are missed [67].

Patients with COPD often adapt to their condition and will tolerate significantly lower levels of blood oxygen without perceived difference to their health or impact on everyday activities. A simple threshold would determine such patients at risk of deterioration.

**Figure 34** shows the daily SpO2 time-series data for a patient with COPD and the wide variation that is typically observed; short term and long term average are shown, and the vertical dotted lines indicate clinical event.

Using a simple threshold would result in both false negative (value below threshold without clinical event) and false positive (value above threshold during clinical event) indications for this patient.

However when the residual (difference between actual and average) is analysed [68], significant variation is observed, and the standard deviation of the residual is seen to increase significantly during periods of exacerbation (**Figure 35**). Physiologically this might be explained by the patient having a coping mechanism for low blood oxygen, but during exacerbation, this mechanism is overwhelmed or fails, and results in the large variations that are observed, giving a clearer indication of impending deterioration.

**Figure 34.** *Daily SpO2 time-series data for patient with COPD [68].*

**Figure 35.** *Residuals (top) and estimates of standard deviation of residuals (bottom) [68].*
