The PRACTICE study was developed using implementation science theory and formative qualitative research. Intervention design was informed by the Consolidated Framework for Implementation Research (CFIR) to identify barriers and facilitators to ICT delivery, while educational and behavioral theories and human-centered design principles guided the development and digitization of the BL implementation package. The resulting package was evaluated in a pragmatic cluster-randomized controlled trial (RCT) [15,16,30].

PRACTICE was a cluster-RCT conducted in the Machinga and Balaka districts of southern Malawi from May 2022 to September 2023. PRACTICE was described in detail previously [16]. Briefly, the study took place in 34 health care facilities, 14 in Balaka and 20 in Machinga, all of which routinely offer HIV testing and treatment services. These facilities included hospitals (n=3), health centers (n=27), and dispensaries (n=4), the latter being smaller facilities that provide basic primary health care. All facilities were operated by either the Malawi Ministry of Health (MOH) or the Christian Health Association of Malawi (CHAM), a faith-based organization. Both the MOH and CHAM provide HIV services free of charge under the national program. The selected sites received additional implementation support from the Tingathe Program, a US President’s Emergency Plan for AIDS Relief (PEPFAR)–supported initiative led by Baylor College of Medicine Children’s Foundation Malawi.

For randomization purposes, the 34 health care facilities were grouped into 33 clusters, with 2 neighboring facilities with similar baseline characteristics combined into a single cluster to allow for evenly sized allocation groups. Clusters were stratified by district, facility type, and the baseline number of potential index clients, then randomized in a 2:1 ratio to receive either SOC or SOC + BL.

ICT is an evidence-based approach used to identify individuals who may have been exposed to HIV. In Malawi, HCWs support persons living with HIV in identifying and testing their sexual partners and family members. This study considered index clients to be persons living with HIV who were either newly diagnosed through HIV testing services (HTS) at a facility or previously diagnosed and receiving antiretroviral therapy (ART). Potential contacts include the index clients’ sexual partners, biological children, household members, and social contacts [16].

At the time of the study, ICT in Malawi was implemented through three referral methods: (1) active or assisted referral, in which an HCW confidentially notifies and traces contacts for testing; (2) passive referral without self-test kits, in which the index client independently encourages contacts to test using a referral slip; and (3) passive referral with HIV self-test kits, in which the index client provides contacts with both a referral slip and an HIV self-test kit. Effective ICT implementation relies on trained lay HCWs to provide counseling and organize appropriate logistics for tracing contacts.

The SOC consisted of 1 centralized, in-person, facilitator-led, 2-hour ICT training for HCWs, covering key ICT concepts, procedures, and counseling approaches. This one-time training was part of a larger Tingathe training that covered a range of areas beyond ICT. The SOC implementation was also supplemented by regular in-person facility mentorship visits conducted across all 33 Tingathe-supported clusters.

As previously described [16,29], clusters randomly assigned to the BL implementation package arm (n=11) received SOC + BL, which combined digital and in-person training with monthly CQI sessions. The intervention was structured into 4 phases. First, all HCWs completed an 8-hour asynchronous digital learning module over weeks 1 to 3 using tablets at their facilities. Following week 3, a 2-day (14-hour), in-person synchronous training was held over a weekend at or near their facility. These small-group, tablet-guided sessions included role-plays, individual feedback, and group discussions focused on ICT protocols. In weeks 4 to 5, all HCWs received a 1-hour phone feedback session based on simulated ICT encounters with study staff portraying index and contact clients; study staff provided feedback, and HCWs who struggled to master the material were given opportunities to repeat the exercises until their performance improved. Finally, from weeks 10 to 52, HCWs participated in a total of six 2-hour facility-based CQI sessions to review implementation, troubleshoot challenges, and refine strategies. In total, the BL components layered onto SOC accounted for approximately 35 additional hours of training, in addition to the 2 hours provided through SOC.

The base case represents the cluster-RCT scenario, modeling the estimated costs and outcomes for index clients enrolled in the 33 clusters participating in the PRACTICE study over 1 year [16,29]. The trial observed approximately 50,000 potential index clients receiving services at the 33 clusters per year. Potential index clients were defined as the sum of individuals enrolled in the ART program and those newly diagnosed through the HTS program over 1 year. To address the asymmetry of the 2:1 randomization, we simulated all 33 clusters receiving the intervention package (50,000 potential index clients) compared with all 33 clusters receiving the SOC (50,000 potential index clients), for a total of 100,000 index clients modeled.

Both study arms received standard ICT training and programming; thus, costs associated with the BL components were treated as incremental to SOC. SOC-related costs were assumed to be equivalent across arms and were not estimated separately.

Consistent with Malawi’s public health system, analyses were conducted from a health system perspective, incorporating system-level costs related to intervention implementation, tracing, and testing contacts of index cases living with HIV. We estimated the total incremental costs associated with developing and implementing BL over 1 year of operation. All costs and effectiveness outcomes were annualized to reflect this 1-year analytic time horizon. Our primary cost-effectiveness outcomes included cost per client tested and cost per person diagnosed with HIV, estimated using incremental cost-effectiveness ratios (ICERs).

Event probabilities were estimated using observed clinical trial outcomes when available and supplemented by clinic register review and published literature. Plausible ranges were derived from reported 95% CIs, the minimum and maximum values found across relevant studies, or, when ranges were not explicitly reported, from the spread of values across comparable published estimates. When no published or empirical reference was available to inform parameter uncertainty, ranges were defined by applying a ±20% variation around the base-case value (Table 1 and Figure 1). The model was constructed to reflect the trajectory of index and contact client engagement with the health system and affiliated HCWs, from the point of index presentation to contact testing and HIV test outcomes.

We included costs related to both developing and implementing the BL implementation package, including all 4 BL components described earlier (Multimedia Appendix 1). Costs were further designated as strategy development or implementation costs (Table 2).

Cost data were obtained through microcosting and time-and-motion assessments. Microcosting relied on program receipts, program accounts, administrative records, and human resource records. Time-and-motion data were gathered through 3 distinct methods: self-reports and timed assessments from HCWs conducting field-based tracing (Multimedia Appendix 2), direct observations by study staff for phone tracing, and tablet-based paradata tracking of the time spent on digital learning tasks. Time estimates from each method were converted to costs using relevant HCW cadre salaries. In cases in which daily subsistence allowances were provided for a specific activity, we used the daily subsistence allowances as the cost input and did not double-count the associated salary. Costs were categorized at the district, facility, and HCW levels to inform scale-up projections (Multimedia Appendix 2).

Costs were calculated in 2023 Malawian Kwacha (MWK) and 2023 US $. The exchange rate used during the bulk of the trial period was US $1=MWK 1036 [33].

We developed a decision tree model to simulate costs, contact tracing and contact testing outcomes, and estimate the ICERs comparing SOC with SOC + BL (Figure 1). Based on annual visit estimates for individuals presenting for HTS or ART care at 33 participating clusters, our base-case model reflects 50,000 potential index clients.

Our decision tree structure reflects the intentional separation between index and contact registers, in which index clients and contact clients are not directly linked to maintain confidentiality. As such, the model is presented in 2 subparts. Figure 1A reflects the movement of presenting index clients, beginning with a decision node (black square) assigning index clients to either the SOC or SOC + BL, using a 50:50 allocation. Within each arm, a chance node (black circle) determines whether the index client was identified via HTS or ART registers. This is followed by another chance node that models whether the individual participates in ICT, defined as naming contacts for tracing and testing.

Named contacts elicited from participating index clients then serve as the basis for the population moving through Figure 1B. The number of contacts elicited per index client was derived from trial data and varied by arm and source (SOC + BL/ART, SOC + BL/HTS, SOC/ART, and SOC/HTS). These estimates were applied externally by multiplying the number of index clients who participated in ICT (from Figure 1A) by the average number of contacts elicited per participating index client.

Once elicited, contact clients are categorized by HIV status as either unknown (requiring HIV testing) or known HIV-positive but not on ART. Contacts with unknown status proceed through 1 of 3 tracing methods according to the referring index client’s preference: active referral, passive referral with self-test kits, or passive referral without self-test kits. A series of chance nodes determine whether the contact is successfully traced, whether they undergo HIV testing, and, if tested, whether they receive a positive or negative result.

We made several key assumptions that influence the model structure and potentially the outcomes. In this 2:1 cluster-RCT, only 11 clusters were exposed to the intervention. In our model, we present outcomes and associated costs as if all 33 clusters were exposed to the intervention, scaling district, facility, and HCW inputs to reflect this assumption. This is then compared with the SOC, again reflecting a scenario in which all 33 clusters received only standard services.

Further, we assumed that the proportion of index clients identified through either HTS or ART was independent of the intervention arm (Table 1). Finally, as the trial did not collect data on the proportion of named contacts who were traced, we used a base-case assumption that all named contacts in the active referral pathway were traced. The trial also did not collect testing rates by notification approach among named contacts; these estimates were obtained retrospectively through manual review of HIV testing registers at participating health facilities. All inputs were varied over a plausible range and tested in univariate and probabilistic sensitivity analyses to examine their relative importance, as described below.

We calculated ICERs by dividing the difference in costs by the difference in each outcome for the SOC + BL ICT implementation approach compared with the SOC approach.

We built the model in Microsoft Excel, version 16.97.2 (25052611), and used Crystal Ball (version 11.1.2.4.900; Oracle) to conduct probabilistic sensitivity analyses across 1000 simulations drawn from defined individual parameter distributions.

To assess the robustness of the model, we conducted both 1-way (univariate) and probabilistic sensitivity analyses. The 1-way sensitivity analyses examined the impact of varying individual parameters on key outcomes to identify which inputs most influenced model results [34]. The probabilistic uncertainty analyses varied all parameters simultaneously based on their distributions, capturing the combined effect of parameter uncertainty on model outcomes [35]. We also conducted a sensitivity analysis of the US $–MWK exchange rate to assess the impact of fluctuations during the study period, which spanned October 2022 to September 2023.

We conducted 2 scenario analyses to explore the cost implications of alternative implementation approaches for the BL implementation package. First, we estimated implementation costs excluding one-time development costs, such as strategy design, content creation, and training materials, to simulate annual implementation across the original 33 clusters. This provided insights into the cost-effectiveness of a realistic implementation scenario in which development costs would not be duplicated, as these costs are otherwise included in the base-case approach.

Second, we modeled a nationwide rollout of the BL implementation package across all facilities providing HIV testing and ART services in Malawi over a 1-year implementation horizon. This scenario reflected a pragmatic, government-led implementation approach incorporating key efficiencies expected under MOH leadership by assuming integration within existing health system structures and personnel tasks. We used 2024 national data on new ART initiations and ongoing ART use to estimate the eligible index testing population, defining “initiations” as persons newly diagnosed with HIV (ie, HTS in our model). To estimate HCW needs, we scaled the number of HCWs per facility by client volume, using the base-case average (n=9) as a reference.

Implementation was envisioned through a decentralized cascade model. A 4-day central training of trainers (TOT) was modeled after existing programs for similar activities. The TOT was intended for district-level leadership, including 1 district health officer, ART coordinator, HTS coordinator, and driver from each district, along with 4 MOH staff members and 1 national trainer. Under this approach, efficiencies were expected through reduced reliance on external facilitators, use of existing MOH training spaces, and integration of ICT duties into the routine responsibilities of district and facility staff. Development costs were excluded, assuming use of already developed materials for national rollout. Cost categories were scaled based on facility, district, and personnel counts, with recurring costs (training delivery, travel, venues, and refreshments) adjusted to reflect district-led implementation. Fuel and venue costs were substantially reduced compared with the base case, consistent with local delivery and MOH-owned or subsidized training spaces. Detailed operational assumptions, including the composition of the TOT cohort, district-level training logistics, per diem and travel policies, and CQI delivery procedures, are provided in Multimedia Appendix 3.

To account for the possibility that training materials may require periodic updating or adaptation over time (eg, in response to evolving guidelines or contextual needs), we examined an additional scenario incorporating partial reuse of content development and video production materials (Multimedia Appendix 4). This scenario reflects reduced development costs associated with maintaining and updating existing materials rather than full redevelopment.

The PRACTICE trial was approved by the Malawi National Health Science Research Committee (#20/06/2566), the University of North Carolina at Chapel Hill Institutional Review Board (20-1810), and the Baylor College of Medicine Institutional Review Board (H-48800). All participating HCWs provided written informed consent prior to engaging in any study-related activities, including training sessions and program implementation.

In the base case, we evaluated 50,000 potential index clients receiving services in the trial-included clinics over 1 year, modeling this cohort of 50,000 clients receiving SOC + BL compared with the same cohort receiving SOC. SOC + BL yielded a total of 9374 named contacts, compared with 7122 named contacts under SOC. Among these, we estimated that 3026 contacts would be tested and 183 individuals would be newly diagnosed with HIV under SOC + BL, compared with 2135 contacts tested and 129 individuals newly diagnosed under SOC. SOC + BL therefore resulted in 891 additional contacts tested and 54 more persons diagnosed with HIV (per 50,000) compared with SOC, representing approximately a 42% relative increase in both contact testing and HIV diagnoses. The total cost of implementing the BL strategy in the 33 clusters across 2 districts was US $111,258, resulting in an ICER of US $125 per additional contact tested and US $2045 per additional person diagnosed with HIV (Table 3).

One-way sensitivity analyses identified the key drivers of uncertainty in the cost-effectiveness model. For cost per contact tested, ICERs were most sensitive to probabilities within the SOC condition, specifically, the probability that index clients identified through ART services would engage in ICT and the probability that index clients identified through HTS (ie, newly diagnosed) would opt for passive referral for their named contacts. Variations in these parameters shifted the estimated ICER for contact testing by up to ±US $34, indicating substantial leverage on cost-effectiveness.

For cost per person diagnosed with HIV, ICERs were again primarily driven by probabilities estimated in the SOC condition, specifically, the probability that an index client identified through HTS would choose passive referral and the probability of ICT enrollment among individuals identified through HTS. These parameters produced the largest ICER swings for this outcome, shifting the estimated cost per person diagnosed with HIV by ±US $850. Together, these findings indicate that both cost per contact tested and cost per person diagnosed with HIV are far more sensitive to programmatic outcomes than to associated implementation costs. Overall, these findings highlight that the robustness of the cost-effectiveness estimates depends on accurate empirical data for referral pathways and ICT uptake.

The full table of 1-way sensitivity analysis results comparing SOC and SOC + BL, along with tornado diagrams showing percentage changes in ICERs across plausible parameter ranges, are provided in Multimedia Appendices 5 and 6.

Figure 2 presents cost-effectiveness planes from the probabilistic sensitivity analysis, showing the joint distribution of incremental costs and effects for SOC compared with SOC + BL. Figure 2A plots incremental cost against the number of contacts tested, while Figure 2B plots incremental cost against the number of persons diagnosed with HIV. Each point represents 1 simulation drawn from the parameter distributions.

For the cost per contact tested, nearly all (966/1000, 96.6%) simulations fell in the northeast quadrant, indicating that the BL implementation package was both more effective and more costly than SOC. For the cost per person diagnosed with HIV, 58% (580/1000) of simulations were in the northeast quadrant, while 42% (420/1000) were in the northwest. These findings suggest that although the BL implementation package strategy typically improves health outcomes, its cost-effectiveness is sensitive to variations in key input parameters.

To assess the impact of exchange rate fluctuations during the study period, we conducted a sensitivity analysis comparing the base-case exchange rate of MWK 1036=US $1 (reflecting November 2023) with a lower exchange rate scenario of MWK 811=US $1 (reflecting November 2021). These time points correspond to the start and end of the PRACTICE study implementation period. Under the MWK 811=US $1 scenario, total program costs increased from US $111,258 to US $141,910, and cost-effectiveness estimates increased from US $125 to US $159 per contact tested and from US $2045 to US $2609 per person diagnosed with HIV. These findings indicate a moderate effect of exchange rate variability on program costs and cost-effectiveness. Multimedia Appendix 7 [36] provides additional context and assumptions.

Removing strategy development costs reduced the total program cost from US $111,258 in the base case to US $61,774 (Table 3). This led to a substantial decrease in the estimated ICER, with the cost per contact tested falling by US $56 (from US $125 to US $69) and the cost per person diagnosed with HIV declining by US $909 (from US $2045 to US $1136).

In the pared-down, government-led, nationwide scale-up scenario, annual program costs increased to US $727,113, while the cost per contact tested decreased to US $43 and the cost per person diagnosed with HIV declined to US $698.

Figure 3 shows the results of the probabilistic sensitivity analysis for the nationwide expansion scenario, using the same ICER plane structure described above. For the cost per contact tested, 96.4% (964/1000) of simulations fell in the northeast quadrant, indicating that SOC + BL was more effective and more costly than SOC. For the cost per person diagnosed with HIV, just over half (552/1000, 55.2%) of simulations were in the northeast quadrant and 44.8% (448/1000) were in the northwest quadrant. No simulations fell in the southeast quadrant, indicating minimal potential for scenarios in which SOC + BL would result in both cost savings and improved outcomes relative to SOC.