Evaluation of 6 years of eHealth data in the alcohol use disorder field indicates improved efficacy of care

Abstract

Background: Predictive eHealth tools will change the field of medicine, however long-term data is scarce. Here, we report findings on data collected over 6 years with an AI-based eHealth system for supporting the treatment of alcohol use disorder.

Methods: Since the deployment of Previct Alcohol, structured data has been archived in a data warehouse, currently comprising 505,641 patient days. The frequencies of relapse and caregiver-patient messaging over time was studied. The effects of both introducing an AI-driven relapse prediction tool and the COVID-19 pandemic were analyzed.

Results: The relapse frequency per patient day among Previct Alcohol users was 0.28 in 2016, 0.22 in 2020 and 0.25 in 2022 with no drastic change during COVID-19. When a relapse was predicted, the actual occurrence of relapse in the days immediately after was found to be above average. Additionally, there was a noticeable increase in caregiver interactions following these predictions. When caregivers were not informed of these predictions, the risk of relapse was found to be higher compared to when the prediction tool was actively being used. The prediction tool decreased the relapse risk by 9% for relapses that were of short duration and by 18% for relapses that lasted more than 3 days.

Conclusions: The eHealth system Previct Alcohol allows for high resolution measurements, enabling precise identifications of relapse patterns and follow up on individual and population-based alcohol use disorder treatment. eHealth relapse prediction aids the caregiver to act timely, which reduces, delays, and shortens relapses.

Keywords: addiction; alcohol; eHealth; prediction; relapse.

Figure 1

The average of relapse rate (RPPD; relapse per patient day) over time since introduction of previct alcohol (blue solid line) with estimated error range (±3*standard error, grey dashed lines). The start of the timeline represents Oct 1st 2015. The COVID-19 pandemic (March 2020–Jan 2022) is indicated as light grey. The predictor was implemented late in May 2021 and the associated period is indicated as light grey. The overlap between the pandemic and implementation is indicated as dark grey.

Figure 2

(A) Caregiver interactions through the eHealth tool (messages sent per day) decays over treatment time from an initial about 0.1 message sent per patient day to about 0.04 messages per patient day after 200 days in treatment. (B) The relative temporal patterns of coinciding events during treatment using cross-correlation are shown. Relative probability of messages sent to the patient (y-axis) at a lag of time (x-axis) after activation of the predictor (at day 0) is shown as blue solid line. The average probability for (any) message per patient day for the period was used for normalization is shown as dashed grey line. Days 1–3 after activation of the predictor, it is 10 times more probable that a message is sent, followed by a decline of messaging probability towards the average level.

Figure 3

Coinciding events during treatment. The relative temporal patterns as investigated using cross-correlation are shown. (A) Relative risk (y-axis) of relapses of durations greater than 0, 2, 4 days observed for patients at a lag of time (x-axis) after activation of the predictor (at day 0) of are shown. Dashed lines indicate values computed from data collected from the “pre” period, prior to implementation of the predictor. Solid lines indicate values computed from data collected from the “post” period, after the implementation of the predictor. Red, Blue, and Cyan lines represent relapses greater than 0, 2, and 4 contiguous days. In the implementation of the predictor, activation cannot occur during a day in relapse. During the first 9 days following predictor activation, the probability of relapse is consistently smaller for “post” period data, indicating that the implementation of the predictor reduced the relapse quantity. (B) The relative reduction of risk for relapse in percent comparing the pre and post periods for duration (y-axis) greater than 0, 1, 2, 3, 4 contiguous days in relapse (x-axis) are shown as bars.