Prioritizing COVID-19 Contact Tracing During a Surge Using Chatbot Technology

Abstract

When COVID-19 cases surge, identifying ways to improve the efficiency of contact tracing and prioritize vulnerable communities for isolation and quarantine support services is critical. During a fall 2020 COVID-19 resurgence in San Francisco, California, prioritization of telephone-based case investigation by zip code and using a chatbot to screen for case participants who needed isolation support reduced the number of case participants who would have been assigned for a telephone interview by 31.5% and likely contributed to 87.5% of Latinx case participants being successfully interviewed. (Am J Public Health. 2022;112(1):43–47. https://doi.org/10.2105/AJPH.2021.306563)

In this era of effective vaccines and increased federal investment, health departments have reduced staffing while also seeking to retain a workforce with the capacity to contact trace COVID-19.¹ However, the rapidity of resurgence and testing delays as well as the low proportion of case participants naming contacts pose major challenges to program effectiveness.^2–5 Ensuring that vulnerable populations receive the ancillary services to safely isolate remains a critical health equity goal. Identifying best practices on how to optimally use staff and technology is necessary to maximize efficiency and ensure the capacity to support isolation.

INTERVENTION

In response to a surge in cases in fall 2020, the San Francisco Department of Public Health (SFDPH), in California, implemented a tiered prioritization strategy that focused telephone-based case investigation on certain zip codes while using a secure, confidential chatbot tool to maintain ongoing outreach to all case participants and screen for those who needed isolation support. We conducted this work as part of SFDPH’s COVID-19 surveillance.

The Virtual Agent is an automated, interactive chatbot integrated with CalCONNECT, California’s statewide case management and contact-tracing platform. Upon health department receipt of the lab result, we invited case participants to complete the chatbot survey through a unique link sent via text message. Available in English and Spanish, the Virtual Agent requires case participants to confirm their date of birth and zip code; it then educates individuals about their isolation period and asks about their ability to isolate, symptoms, spread settings, affiliation with congregate settings, and demographics. The Virtual Agent also elicits close contacts, thus allowing contact notification to occur without needing to conduct a telephone interview with the case participant (Appendix A [available as a supplement to the online version of this article at http://www.ajph.org]). We saved the data the Virtual Agent collected to a MongoDB hosted in Amazon Web Services DocumentDB (Seattle, WA). This is a custom solution built by Accenture exclusively for the California Department of Public Health. Government agencies can contact the California Department of Public Health and Accenture (Samantha.Geronimo@accenture.com and Eyal.Darmon@accenture.com) about how to use this.

PLACE AND TIME

To examine this approach, we compared COVID-19 case participant interview outcomes in San Francisco by race/ethnicity during the two-month surge period (November 16, 2020–January 15, 2021) to the outcomes of the previous two, “nonsurge” months (September 1, 2020–October 31, 2020).

PERSON

This program included all COVID-19 case participants reported to the SFDPH who were assigned for case investigation. We excluded case participants with test results received 10 or more days after specimen collection and those residing in congregate living settings.

PURPOSE

Beginning in November 2020, there was a steep increase in the COVID-19 case rate, leading to more case participants identified over a two-month period than in the previous seven months combined. This surge exceeded the SFDPH’s capacity to provide traditional telephone-based case investigation for all case participants, so the SFDPH focused its intensive telephone-based efforts on the zip codes with the highest case rates, which correlated with where Latinx communities are known to reside. In San Francisco, Latinx individuals represent 15% of the population but made up 40% of identified COVID-19 cases, and they are more likely to request support to isolate and have contacts test positive than any other race/ethnicity.^6,7

IMPLEMENTATION

In November 2020, the SFDPH began sending the chatbot survey to all case participants assigned for case investigation (except those younger than 18 years because of lack of informed consent) and implemented a two-tiered prioritization schema for telephone calls. Factors informing inclusion in the high-priority tier 1 included age (younger than 18 years and 50 years and older) and residence in a priority zip code (top six zip codes based on current case rates and data regarding proportion of case participants successfully completing an interview and naming contacts). We sent tier 1 case participants the chatbot survey and made at least two telephone call attempts over two days regardless of chatbot response. We sent tier 2 case participants the chatbot survey only. In both tiers, we immediately called all chatbot respondents who indicated an inability to safely self-isolate, interviewed them, and offered them support services.

We considered an interview complete if the case participant completed or partially completed the telephone interview or responded to the chatbot and answered at least the first question regarding their ability to isolate.

We created a data set composed of cases in SQL Server Management Studio (Microsoft Corp., Redmond, WA) from public health databases and analyzed it using R software (R Foundation for Statistical Computing, Vienna, Austria). We tested statistical differences in group outcomes using the Fisher exact test.

EVALUATION

The number of case participants assigned for case investigation increased by 416% in the nonsurge compared to the surge period (2731/14 095), whereas case investigation staffing increased only 20% (64.2 vs 77.0 full-time equivalent; Table 1). During the surge period, 68.5% (9656/14 095) of case participants met tier 1 criteria, and we prioritized them for a telephone interview. We did not assign the 31.5% (4439/14 095) who met tier 2 criteria for telephone interview.

TABLE 1—

COVID-19 Case Investigation and Contact-Tracing Outcomes: San Francisco, CA, September 1–October 31, 2020, and November 16, 2020–January 15, 2021

	Nonsurge	Surge
All, No. (Column % or % of Eligible)	All, No. (Column % or % of Eligible)	Tier 1, No. (Row % or % of Eligible)	Tier 2, No. (Row % or % of Eligible)
Eligible case participants
Total	2 731 (100.0)	14 095 (100.0)	9 656 (68.5)	4 439 (31.5)
Race/ethnicity
White	544 (19.9)	3 115 (22.1)	1 471 (47.2)	1 644 (52.8)
Latinx	1 200 (43.9)	5 339 (37.9)	4 522 (84.7)	817 (15.3)
Other (group)a	987 (36.1)	5 641 (40.0)	3 663 (64.9)	1 978 (35.1)
Asian	439 (16.1)	2 512 (17.8)	1 764 (70.2)	748 (29.8)
Black or African American	136 (5.0)	698 (5.0)	545 (78.1)	153 (21.9)
Multiracial	83 (3.0)	146 (1.0)	101 (69.2)	45 (30.8)
Native American	3 (0.1)	37 (0.3)	29 (78.4)	8 (21.6)
Native Hawaiian/Pacific Islander	30 (1.1)	138 (1.0)	115 (83.3)	23 (16.7)
Other	65 (2.4)	1 164 (8.3)	744 (63.9)	420 (36.1)
Declined to state	13 (0.5)	. . .	. . .	. . .
Unknown	218 (8.0)	946 (6.7)	365 (38.6)	581 (61.4)
Case participants interviewed b
Total	2 335 (85.5)	10 664 (75.7)	7 915 (82.0)	2 749 (61.9)
Race/ethnicity
White	475 (87.3)	2 289 (73.5)	1 155 (78.5)	1 134 (69.0)
Latinx	1 122 (93.5)	4 672 (87.5)	4 083 (90.3)	589 (72.1)
Other (group)a	738 (74.8)	3 703 (65.6)	2 677 (73.1)	1 026 (51.9)
Case participants interviewed in 24 h
Total	2 099 (76.9)	7 216 (51.2)	5 023 (52.0)	2 193 (49.4)
Race/ethnicity
White	436 (80.1)	1 748 (56.1)	809 (55.0)	939 (57.1)
Latinx	1 015 (84.6)	3 022 (56.6)	2 573 (56.9)	449 (55.0)
Other (group)a	648 (65.7)	2 446 (43.4)	1 641 (44.8)	805 (40.7)
Case participants who named ≥ 1 contacts
Total	1 244 (45.6)	3 629 (25.7)	3 158 (32.7)	471 (10.6)
Race/ethnicity
White	250 (46.0)	553 (17.7)	394 (26.8)	159 (9.7)
Latinx	610 (50.8)	1 877 (35.2)	1 733 (38.3)	144 (17.6)
Other (group)a	384 (38.9)	1 199 (21.2)	1 031 (28.1)	168 (8.5)
Mean no. contacts named
Total	1.35	0.74	0.96	0.27
Race/ethnicity
White	1.27	0.46	0.68	0.27
Latinx	1.58	1.03	1.13	0.43
Other (group)a	1.13	0.63	0.86	0.21
Eligible contacts elicited	3 121	9 660	7 804	1 856
Contacts notified	2 639 (84.6)	7 314 (75.7)	6 215 (79.6)	1 099 (59.2)
Investigators	64.2	77	. . .	. . .
Average cases per investigator	42.5	183.1	125.4	. . .

Note. We stratified results by race and surge period. The nonsurge period was September 1, 2020–October 31, 2020. The surge period was November 16, 2020–January 15, 2021.

^a “Other (group)” combines those responding as Asian, Black or African American, Multiracial, Native American, Native Hawaiian/Pacific Islander, other, declined to state, and unknown.

^b We defined interview completion by a completed or partially completed telephone interview or a response to the chatbot in which at least the first question regarding the case participant’s ability to isolate was answered.

Across both tiers, we successfully sent 90.5% (11 303/12 491) of case participants aged 18 years and older the chatbot survey, and among those 37.3% (4221/11 303) responded (data not shown). Chatbot response rates were higher among White case participants than among Latinx case participants (53.6% [1423/2656] vs 31.4% [1300/4140]; P < .001), and higher among those aged 18 to 49 years than among those aged 50 years or older (41.3% [3501/8480] vs 25.5% [720/2823]; P < .001; data not shown).

Although the overall proportion of case participants interviewed decreased from 85.5% (2335/2731) to 75.7% (10 664/14 095) between periods, the percentage of Latinx case participants who completed interviews remained high (93.5% [1122/1200] during nonsurge vs 87.5% [4672/5339] during surge; Table 1). Among completed interviews, we conducted 21.0% (2236/10 664) using exclusively the chatbot, with a higher proportion of White case participants interviewed with the chatbot only compared with Latinx case participants (42.6% [976/2289] vs 8.9% [414/4672]; data not shown). The proportion of case participants interviewed within 24 hours of test receipt decreased from 76.9% (2099/2731) to 51.2% (7216/14 095) across periods. The median time from when the chatbot was sent to a response was 26 minutes.

The number of contacts identified across periods increased 210%, from 3121 to 9660, whereas the mean number of contacts named per case participant decreased from 1.35 to 0.74 (Table 1). The number of contacts named per case participant was similar between White and Latinx case participants during nonsurge (1.27 vs 1.58) but decreased during surge (0.46 vs 1.03).

During the nonsurge period, the mean case load per case investigator per month was 42.5, which increased to 183.1 during surge (Table 1). However, when considering only tier 1 case participants assigned for telephone interview, the mean case load per case investigator during the surge was 125.4, which is 31.5% lower than the total mean case load of 183.1.

ADVERSE EFFECTS

Telephone interviews were more likely than the chatbot to yield named contacts; however, chatbot technology helped ensure that all individuals were offered isolation support services within minutes of receiving test results.

SUSTAINABILITY

Prioritization of telephone outreach by zip code and using a chatbot reduced the number of case participants who would have been assigned for telephone interview by 31.5% and likely contributed to 87.5% of Latinx case participants being successfully interviewed and more than half interviewed within a day of test results. This high percentage of completed interviews among a population at higher risk was achieved despite the number of cases increasing several fold between the nonsurge and surge periods. Given planned reductions in contact-tracing staffing, this strategy will be sustained to address future surges and disease outbreaks.

The zip code–based prioritization strategy may not be generalizable to jurisdictions where zip codes do not align as much with race/ethnicity, income, and education level. In San Francisco, more than 85% of Latinx case participants reported to the SFDPH resided in a priority zip code.

PUBLIC HEALTH SIGNIFICANCE

Given competing demands on health departments during a COVID-19 resurgence, programs could experiment with more automated models of offering timely isolation resources to vulnerable communities. Systematically integrating chatbot technology into disease investigation may improve timely outreach and modernize health departments in preparation for future disease outbreaks.