Data inputs for HIV models and how they inform HIV prevention interventions Dobromir Dimitrov Vaccine & Infectious Disease Division Fred Hutchinson Cancer Research Center
Dynamic model integration The mechanisms used to integrate processes at different layers are critical for the structural soundness of HIV models. In the core (bottom two layers) is the structure of the population when unexposed to infection. They handle demographic processes and sexual behavior. Disease transmission under the current standard of care (in absence of the intervention) is modeled at the epidemic layer. The intervention layer handles mechanisms of reducing susceptibility and/or infectiousness and other specifics of implementation.
Typical stages of a modelling process Assessment of situation: question, epidemiological setting and data available à to inform model structure required Model development and coding Data analysis or reviews to inform pre-intervention parameters and fit the model to data Design intervention scenarios (specify intervention parameters) Multiple simulations (1000 s) for each scenarios Analysis Results Feed-back
Model inputs Case 1: Informing preintervention parameters (demography, sexual behavior, HIV transmission, etc.) Goal: Closely represent the epidemic conditions Sources: published literature, national and regional surveys, clinical trial questionnaires Challenges: missing data, multiple sources, different methodology Addressed by: model calibration Case 2: Informing intervention parameters (delivery, efficacy profile, coverage, adherence, etc.) Goal: Optimize the potential benefits from future interventions Sources: clinical trial results, expert opinion, targets Challenges: missing data, knowledge gaps, uncertainty in mechanisms of protection Addressed by: sensitivity analyses
Case 1: Modeling HPTN 078 Enhancing Recruitment, Linkage to Care and Treatment for HIV-Infected Men Who Have Sex with Men (MSM) in the United States https://www.hptn.org/research/studies/154 Intervention (n = 178) Case Manager Intervention Package MSM Not virally suppressed Control (n = 178) SOC for Linkage and Treatment Deep-Chain Respondent Driven Sampling Individual Randomization Modelling: predict population-level HIV incidence reduction
Case 1: Model structure HIV disease progression: CD4 decline by viral load Risk groups: age (<25, 25+) x race (black, white) Care cascade: On treatment Never testing, undiagnosed Testing, undiagnosed Diagnosed In care Achieving suppression Adherent Non-adherent Dropped out Modeling study of the HPTN Modeling Center with lead author Dr. Kate Mitchell from Imperial College.
Case 1: Model inputs Domain Examples Data source Disease progression Infection probabilities Efficacy of HIV prevention and treatment Sexual risk behaviour Initial CD4 and viral load distribution HIV-related mortality CD4 progression rates Per-sex-act transmission probability Relative infectiousness different disease stages Reduction in HIV transmission risk: condoms, ART Number and type of partners Condom use Age and race of partners Care cascare HIV testing Linkage/dropout from HIV care ART linkage and dropout Published studies: cohorts in North America and Europe Published studies: meta analyses, study of Australian MSM Published studies: clinical trials, metaanalyses NHBS surveillance data, 078 trial NHBS surveillance data, clinical cohorts, state health department data, 078 trial
Case 1: Model calibration Ensure that model reproduce observed data on: HIV prevalence - by age and race (NHBS) MSM demography - by age and race (NHBS, census) Viral suppression level (NHBS, Maryland Health Dept, CDC national estimates)
Case 1: calibration challenges Challenge: differences between data sources Age distribution of the MSM population % virally suppressed Dealt with: fitted separately to 2 different demography scenarios (NHBS, census) 2 sources of care cascade data (NHBS, Maryland Health Dept)
Case 1: Results To reduce HIV incidence 20% by demography scenario Absolute increase in viral suppression Absolute increase in % virally suppressed 0 5 10 15 20 Demography data fitted to NHBS census Intervention Target: Reduce HIV incidence by 20% over 2, 5, 10 years 2yrs 5yrs 10yrs Time over which target met Outcome metric: Absolute increase in viral suppression needed to meet the target Results are not affected by uncertainty in the age distribution of the MSM population
Case 1: Results To reduce HIV incidence 20% by demography scenario Absolute increase in viral suppression To reduce HIV incidence 20% - by care cascade data source Absolute increase in viral suppression Absolute increase in % virally suppressed 0 5 10 15 20 Demography data fitted to NHBS census Absolute increase in % virally suppressed 0 5 10 15 20 Care cascade data fitted to NHBS health dept 2yrs 5yrs 10yrs 2yrs 5yrs 10yrs Time over which target met Time over which target met Results are not affected by uncertainty in the age distribution of the MSM population Results are influenced by initial level of viral suppression
Case 1: Summary Mathematical models of HIV prevention present substantial complexity which aggregate multiple assumptions at different layers of integration Pre-intervention parameters are informed by various sources including surveys, data from clinical trials, official statistics, census data which often disagree Despite discrepancies in input data, modeling results may be robust to uncertainty in some parameters while being sensitive to uncertainty in other. This should be taken under account when modeling analysis are used in decision making In the presented HPTN 078 model the results are robust to uncertainty in MSM demography but influenced by uncertainty in levels of viral suppression at the start of the intervention
Case 2: Modeling HIV vaccination Large ongoing clinical trial (HVTN 702) testing multi-dose HIV vaccine 702 V1V2 Response Month 0 1 3 6 6.5 12 12.5 18 24
Case 2: Modeling HIV vaccination Large ongoing clinical trial (HVTN 702) testing multi-dose HIV vaccine Questions on intervention parameters such as: Vaccine efficacy and coverage Vaccination strategies Adherence to multi-dose regimen and revaccination boosters Background epidemic conditions Modeling study of the HVTN Modeling Team, funded by Gates Foundation, with lead author Dr. Simon de Montigny from University of Montreal.
3.1 Case 2: Efficacy, coverage, delivery Vaccine efficacy (VE) Average duration Minimal VE scenario High VE scenario Phase 1 6 months 20% VE 45% VE Phase 2 12 months 80% VE 95% VE Phase 3 6 months 20% VE 45% VE Overall 24 months 50% VE 70% VE HVTN 702 target Targeted coverage 20% and 50% Roll-out strategies Continuous vaccination (constant vaccination rate) Campaign vaccination (mass vaccination every 2 years)
3.2 Case 2: Efficacy, coverage, delivery Effectiveness metric: Cumulative % infections prevented Higher VE is ~50% more effective Efficiency metric: # infections prevented / 1000 vaccinations Higher VE is ~30% more efficient Simulation over 2017-2027 period Continuous vaccination / Campaign vaccination Boxes represent 90% uncertainty interval based on 1000 epidemic conditions
3.2 Case 2: Efficacy, coverage, delivery Effectiveness metric: Cumulative % infections prevented Efficiency metric: # infections prevented / 1000 vaccinations Higher coverage is ~100% more effective Higher coverage is ~23% less efficient Simulation over 2017-2027 period Continuous vaccination / Campaign vaccination Boxes represent 90% uncertainty interval based on 1000 epidemic conditions
5.1 Case 2: Epidemic conditions Assumption of HIV vaccine licensure in 2027 Compare vaccine impact under more or less optimistic epidemic conditions up to Calibrated HIV incidence curves and projections up to licensure Calibrated ART coverage curves and projections up to licensure Projected curves display large uncertainty in the epidemic conditions at the start of the vaccination in 2027
Case 2: Revaccination (boosting) Assumptions related to boosting: Explore two profiles averaging 30% and 60% efficacy over 2 years Explore scenarios with 0% and 20% attrition between boosters New (naive) vaccinees are added to maintain 50% vaccine coverage Individuals who miss 2+ consecutive single-dose booster would need to get the multidose regimen VE profiles of single-dose booster with high and low overall efficacy
Case 2: Effects of boosting and epidemic assumptions Effectiveness metric 2027-2047 : % infections prevented Vaccine is~14% more effective in low incidence Pessimistic (high incidence) setting Optimistic (low incidence) setting
Case 2: Effects of boosting and epidemic assumptions Effectiveness metric 2027-2047 : % infections prevented Vaccine is~14% more effective in low incidence Efficiency metric 2027-2047: # infections prevented / 1000 vaccinations Vaccine isat least 3-fold more efficient in high incidence setting Low incidence setting Pessimistic (high incidence) setting Optimistic (low incidence) setting Revaccination with high efficacy booster low efficacy booster
4.4 Case 2: Summary Different combinations of intervention parameters may and should be explored in search of optimal implementation strategies Assumptions regarding intervention parameters have to be well understood before interpreting modeling results Presented HVTN 702 model shows that the influence of some parameters on the intervention results may depend on the outcome of interest (evaluation metric) used in the analysis. The impact of vaccination coverage is a good example.
4.4 Case 2: Summary HIV incidence at the start of the vaccination may have little influence on the intervention effectiveness but may have significant impact on the intervention efficiency and from there on the cost-effectiveness of the vaccination If periodic boosting is needed, the efficacy of the booster will be of critical importance of the long-term effectiveness of the vaccination.
4.4 Thank you! HPTN Modeling Center: Marie-Claude Boily, Kate Mitchell, Daniel Wood HPTN Statistical Center: Deborah Donnell, Jim Hughes University of Montreal: Simon de Montigny, Benoit Masse HVTN Statistical Center: Peter Gilbert, Zoe Moodie