Introduction: Telemonitoring (TM) of home NIV patients is increasingly used in clinical practice but data to support clinical or cost effectiveness are limited. We aimed to assess the reliability of TM to identify adequate clinical response to home NIV.

Methods: We used a retrospective monocentric database of 100 patients established on home NIV for whom daily TM data were available over a 90-days period and arterial blood gas results. Performances of 11 algorithms made by NIV experts regarding adequate ventilation (defined as: daily NIV use ? 4hrs/day, improved PaCO2, residual events < 10/h and minimal side effects) were evaluated as well as an algorithm determined by unsupervised machine learning.

Results: Patients disease groups were: COPD (48%), obesity hypoventilation syndrome (30%), neuromuscular/chest wall disease (22%). Mean duration of home NIV was 1.9 years, PaCO2 5.95kPa and adherence 8.4h/day. Only 34/100 patients met the predefined criteria for an adequate clinical response. Sensitivity of experts? algorithms to identify a poor clinical response varied from 28 to 97%, specificity varied from 6 to 94%. The percentage of patients correctly classified with the expert algorithm varied from 51 to 71%. The machine learning had a sensitivity of 63%, a specificity of 76% to identify poor clinical response. The model included mean adherence, residual event entropy and the percentage of triggered breaths.

Conclusion: In addition to increasing the cost of delivery of home NIV, TM data appears to be of little use to identify poor clinical response even with the use of advanced machine learning. Larger datasets to allow internal and external validation are needed to establish more effective algorithms.