Understanding Heavy Drinking at Night through Smartphone Sensing and Active Human Engagement

Abstract

Heavy alcohol consumption can lead to many severe consequences. In this paper, we study the phenomenon of heavy drinking at night (4+ drinks for women or 5+ for men on a single evening), using a smartphone sensing dataset depicting about nightlife and drinking behaviors for 240 young adult participants. Our work has three contributions. First, we segment nights into moving and static episodes as anchors to aggregate mobile sensing features. Second, we show that young adults tend to be more mobile, have more activities, and attend more crowded areas outside home on heavy drinking nights compared to other nights. Third, we develop a machine learning framework to classify a given weekend night as involving heavy or non-heavy drinking, comparing automatically captured sensor features versus manually contributed contextual cues and images provided over the course of the night. Results show that a fully automatic approach with phone sensors results in an accuracy of 71%. In contrast, manual input of context of drinking events results in an accuracy of 70%; and visual features of manually contributed images produce an accuracy of 72%. This suggests that automatic sensing is a competitive approach. CCS CONCEPTS• Human-centered computing → Ubiquitous and mobile computing; Ubiquitous and mobile computing design and evaluation methods.

Examining the Social Context of Alcohol Drinking in Young Adults with Smartphone Sensing

“…This is not the intention of this work, as it would oversimplify the problem. However, inferring the social context of drinking would assist or complement other inferences such as drinking occasions, drinking nights, drink vs. drunk in ubicomp and alcohol research [5,6,27,36,86,87,92]. In this respect, the inference of social context might help to provide meaningful and context-aware interventions that might decrease the amounts consumed, and, as a consequence, less adverse alcohol-related consequences.…”

“…The chosen sensing modalities were Accelerometer (ACC), Applications (APP), Location (LOC), Screen (SCR), Battery (BAT), Bluetooth (BLU), Wifi (WIF), and Proximity (PRO). A dataset summary is given in Table 1 and an extensive description is given in [86,92].…”

“…Prior studies that used this dataset primarily considered user-night as the point of analysis (e.g., inferring nights of alcohol consumption vs. no alcohol consumption [92], inferring heavy-drinking nights [86], etc.). However, in this study, we consider drink-level data, that is more fine-grained.…”

“…Values from all three axes of the sensor were collected, 10 seconds continuously, at a frequency of 50Hz, during every minute. we calculated (a) basic statistics from raw sensor data from the X, Y, and Z-axes [92]; (b) aggregated statistics related to acceleration (m, mNew, dm) and signal magnitude area (mSMA) by combining data from three axes [41,69,86]; and (c) angle between acceleration and the gravity vector [86,92]. -Raw (15) {min., max., med., avg., std.}…”

“…based on the categorization provided in google play store [33]. Using the categories and running apps, statistics such as the total number of running apps and the number of running apps based on categories were calculated [62,86,92]. -Count (2) app count, app record -Category (33) normalized 33-bin histogram of 33 application categories Proximity Several statistical featured were derived using raw values of the proximity sensor.…”