Table 1.

The top 10 predictive rhythm features and the associated mean feature weights for two different subtypes for symptom Depressed.

Subtype 0	Weight	Subtype 1	Weight
#missed_calls$⨂$16-hour_PSD$⨂$12-day_window	0.046	light$⨂$amplitude$⨂$2-day_window	0.058
#SMS_sent$⨂$32-hour_PSD$⨂$14-day_window	0.042	light$⨂$amplitude$⨂$14-day_window	0.057
conversation_length$⨂$20-hour_PSD$⨂$14-day_window	0.041	light$⨂$amplitude$⨂$12-day_window	0.057
#incoming_calls$⨂$36-hour_PSD$⨂$14-day_window	0.038	light$⨂$amplitude$⨂$8-day_window	0.057
screen_on_time$⨂$32-hour_PSD$⨂$8-day_window	0.036	light$⨂$amplitude$⨂$10-day_window	0.057
screen_on_time$⨂$32-hour_PSD$⨂$14-day_window	0.036	light$⨂$amplitude$⨂$6-day_window	0.056
screen_on_time$⨂$mean_deviation$⨂$2-day_window	0.032	light$⨂$amplitude$⨂$4-day_window	0.056
#SMS_read$⨂$10-hour_PSD$⨂$4-day_window	− 0.032	screen_on_time$⨂$amplitude$⨂$6-day_window	0.051
#outgoing_calls$⨂$16-hour_PSD$⨂$10-day_window	0.031	screen_on_time$⨂$amplitude$⨂$8-day_window	0.051
screen_on_time$⨂$median_deviation$⨂$6-day_window	0.030	screen_on_time$⨂$amplitude$⨂$4-day_window	0.050

The naming of the features follows the format [Modality]$⨂$[Rhythm Metric]$⨂$[Window Length], which denotes the modality of the sensor data, the rhythm metric, and the window length used for extracting the feature.