Use of molecular HIV surveillance data and predictive modeling to prioritize persons for transmission-reduction interventions

Abstract

Background:

To develop a predictive model to prioritize persons with a transmissible HIV viral load for transmission-reduction interventions.

Methods:

New York City (NYC) HIV molecular surveillance data from 2010–2013 were used to build a model to predict the probability that the partial pol gene of the virus of a person with a transmissible HIV viral load (>1,500 copies/mL) would be genetically similar to that of a person with a new HIV infection (diagnosis at stage 0 or 1 according to the revised Centers for Disease Control and Prevention classification system). Data from 2013–2016 were then used to validate the model and compare it with five other selection strategies that can be used to prioritize persons for transmission-reduction interventions.

Results:

A total of 10,609 persons living with HIV (PLWH) were included in the development dataset, and 8,257 were included in the validation dataset. Among the six selection strategies, the predictive model had the highest area under the receiver operating characteristic curve (AUC) (0.86, 95% confidence interval [CI]: 0.84, 0.88), followed by the “Young men who have sex with men (MSM)” (0.79, 95% CI: 0.77, 0.82), “MSM with high viral loads” (0.74, 95% CI: 0.72, 0.76), “Random sample of MSM” (0.73, 95% CI: 0.71, 0.76), “Persons with high viral loads” (0.56, 95% CI: 0.54, 0.59), and “Random sample” (0.50, 95% CI: 0.48, 0.53) strategies.

Conclusions:

Jurisdictions should consider applying predictive modeling to prioritize persons with a transmissible viral load for transmission-reduction interventions and to evaluate its feasibility and effectiveness.

In the first two decades of the HIV/AIDS epidemic in the United States, HIV prevention programs primarily focused their efforts on HIV-negative persons at high risk for infection.^[1] Focusing only on persons who are HIV-negative undermines the effectiveness of HIV prevention, as it overlooks the population that is the source of onward transmission—HIV-positive individuals. To maximize reductions in HIV transmission, the Centers for Disease Control and Prevention (CDC) in 2003 recommended incorporating HIV prevention services into the medical care of persons living with HIV (PLWH).^[2]

Lower plasma HIV viral load is associated with lower risk of HIV transmission.^{[3, 4]} In 2011, the landmark HIV Prevention Trials Network (HPTN) 052 study among sero-discordant couples, with the majority (97%) being heterosexual, reported that early initiation of antiretroviral treatment (ART) by the HIV-positive partner reduced the risk of transmission to the HIV-negative partner by 96% compared to delayed treatment.^[5] Two follow-up studies, the PARTNER (Partners of People on ART—A New Evaluation of the Risks) study among heterosexuals and men who have sex with men (MSM) and the Opposites Attract study among MSM, have found zero transmissions between sero-discordant couples when the HIV-positive partner was on treatment and had an undetectable viral load.^{[6, 7]}

The findings from these studies support the strategy of treatment as prevention, i.e., that treating PLWH with ART to prevent HIV transmission be included as a key component of HIV prevention programs. The Panel on Antiretroviral Guidelines for Adults and Adolescents now recommends immediate initiation of ART for all people living with HIV, regardless of CD4 count.^[8] Despite these recommendations, some patients do not initiate ART due to a host of individual- and structural-level factors, whereas others may take ART but are unable to achieve an undetectable viral load due to non-adherence or drug resistance, putting them at risk of transmitting HIV to their negative partners.^[9–11]

One approach to reducing HIV transmission among PLWH, is to identify persons with a transmissible viral load and assist them to achieve viral suppression. This approach is constrained by limited resources, because the number of PLWH with a transmissible viral load at any given time is usually larger than a state or local HIV program can manage. Therefore, a prioritization strategy is needed to identify those at the highest risk of transmitting HIV. Programs have already preferentially selected some sub-populations for intervention, including MSM, Black and Latino people, and persons with high HIV viral loads, including those with acute HIV infection, co-infected with sexually transmitted infections (STIs), or belonging to a “recent and rapid” transmission cluster,^{[3, 12–14]} but there is no systematic way to select individuals for intervention while simultaneously considering multiple factors, e.g., race/ethnicity, transmission risk, age, and viral load, in order to improve the specificity of the targeting strategy and the effectiveness of the intervention.

In the United States, CDC supports local jurisdictions to conduct Molecular HIV Surveillance (MHS), which collects, reports, and analyzes HIV genetic sequences generated during HIV drug resistance testing. It has been used as a tool to identify and respond to PLWH who have one or more viral genetic connections within networks containing recent HIV diagnoses.^[15] The aim of this analysis is to use MHS data and predictive modeling to demonstrate a method that can be used to prioritize PLWH for intervention to reduce HIV transmission.

METHODS

Data source

The data source was the New York City (NYC) HIV surveillance registry. AIDS diagnoses have been reportable in New York State since 1981, and HIV diagnoses have been reportable since 2000. All CD4 counts, viral loads, and nucleotide sequences obtained for genotypic analyses have been reported to the registry since June 1, 2005. As of December 31, 2017, the registry contained a cumulative total of more than 240,000 cases (both living and deceased) and more than 10 million laboratory tests. In 2017, 2,157 people were diagnosed with HIV in NYC and there were about 90,500 PLWH, of whom 7% did not know their HIV-positive status.^[16]

Analysis population

Separating data into development and validation datasets is a common initial step when building a predictive model.^[17] For this analysis, PLWH who had HIV sequence data in the NYC registry and were 13 years of age or older and living in NYC at the end of 2010 with a transmissible viral load, defined as >1,500 copies/mL, were included in the development dataset; PLWH who had HIV sequence data and were 13 years of age or older and living in NYC at the end of 2013 with a transmissible viral load were included in the validation dataset.^[18] Patients who met the following criteria were included in both development and validation datasets: 1) 13 years of age or older by December 31, 2010, 2) diagnosed with HIV by December 31, 2010, 3) alive by December 31, 2013, and 4) viral load >1,500 copies/mL at both times: the end of 2010 and the end of 2013.

Definition of new HIV infection

CDC classifies HIV diagnoses by stage based on patient’s CD4 count at diagnosis and AIDS-defining opportunistic illness. Early infection, defined as a documented negative HIV test within 6 months prior to diagnosis, is classified as stage 0, regardless of CD4 count; CD4 count ≥500 cells/mm³ is classified as stage 1; CD4 count between 200 and 499 cells/mm³ is classified as stage 2; and CD4 count <200 cells/mm³ or an AIDS-defining opportunistic illness (e.g., Kaposi’s sarcoma, pneumocystis pneumonia, and tuberculosis) regardless of CD4 count, is classified as stage 3.^{[19, 20]} Using patient’s stage information, we defined an individual to have a new HIV infection if he/she acquired HIV through a non-perinatal route and was diagnosed with a stage 0 or 1.

Sequence analysis

To determine whether the partial pol sequence of a PLWH was genetically similar to that of a person with a new HIV infection, we used a CDC funded computational tool, Secure HIV-TRACE (HIV TRAnsmission Cluster Engine), following a procedure described previously.^{[14, 21–23]}

When we ran HIV-TRACE on the development dataset to determine whether the partial pol sequence of a PLWH at the end of 2010 was linked, i.e., genetically similar, to that of at least one person with a new HIV infection in NYC in 2011–2013, we included the last sequence from each PLWH at the end of 2010 and the first sequence from each HIV case diagnosed in 2011–2013. First, all sequences were aligned to the HXB2 reference sequence (coordinates: 2253–3869) using an extension of the Smith-Waterman algorithm.^[24] Next, HIV-TRACE calculated the pairwise Tamura-Nei 93 (TN93) genetic distance among all sequences, using an ambiguity fraction of 0.015 (i.e., genetic distance between ambiguous nucleotides were resolved only when the sequence contained ≤1.5% ambiguous nucleotides).^[25] A viral genetic distance ≤0.015 substitution/site between a PLWH and a new infection was considered evidence of similarity, i.e., a link.

When we ran HIV-TRACE on the validation dataset to determine whether the viral sequence of a PLWH at the end of 2013 was genetically linked to that of at least one person with a new HIV infection in NYC in 2014–2016, the same procedure was followed but with different sequence data—the last sequence from each PLWH at the end of 2013 and the first sequence from each HIV case diagnosed in 2014–2016.

Outcome variable

In the development dataset, we included an outcome variable indicating whether a patient’s viral sequence was genetically linked to that of at least one new HIV infection diagnosed in NYC in the next three calendar years, i.e., 2011–2013; in the validation dataset, we included an outcome variable indicating whether a patient’s virus was genetically linked to that of at least one new infection in NYC in the next three calendar years, i.e., 2014–2016.

Model development and validation

We developed our predictive model following the guidelines for Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD),^[26] and using the multivariate adaptive regression splines (MARS) algorithm in the Classification And Regression Training (CARET) package in R.^[27] The candidate variables included sex, transgender status, race/ethnicity, age at diagnosis, current age, transmission category, year of diagnosis, ever diagnosed with AIDS, nadir CD4 count in the last three years, the highest log₁₀ viral load in the last three years, and the log₁₀ last viral load. The final model from the development dataset included the following variables: year of diagnosis, current age, ever diagnosed with AIDS, age at diagnosis, the highest log₁₀ viral load in the last three years, and nadir CD4 count in the last three years.

The predictive model was tested with the validation dataset and compared with five other selection strategies that can be used to select PLWH with a transmissible viral load for transmission-reduction interventions: 1) random sample, 2) persons with high viral loads (sorting PLWHs’ last viral load in descending order and selecting patients with the highest viral loads), 3) random sample of MSM, 4) MSM with high viral loads (sorting MSMs’ last viral load in descending order and selecting patients with the highest viral loads), and 5) young MSM (sorting MSM by age in ascending order and selecting the youngest patients).

In addition to the area under the receiver operating characteristic curve (AUC) to assess the performance of these six selection strategies, we introduced a new measure, the viral genetic linkage rate, which was defined as the proportion of PLWH whose partial pol sequences were genetically linked to that of at least one new HIV infection diagnosed in the next three calendar years. The reason for introducing this new measure is that the performance of each selection strategy also depends on the number of PLWH selected for the transmission-reduction interventions. Using the difference-in-difference analysis, we compared the viral genetic linkage rates by arbitrarily assuming that 250, 500, 750, and 1,000 PLWH, respectively, had been selected for the interventions.

Sensitivity analysis

To assess the sensitivity of our findings, we repeated the above model-building process and analysis by changing the genetic distance threshold from 0.015 substitution/site to 0.005 substitution/site and limiting new diagnoses to acute HIV infections only.

RESULTS

Of 10,609 PLWH living with a documented transmissible viral load in NYC at the end of 2010 and included in development dataset, two-thirds (66.6%) were men and one-third (33.4%) were women; over half (52.2%) were Black and one-third (35.8%) were Hispanic; 5.1% had viruses genetically linked to that of at least one new HIV infection. Of 8,257 PLWH living with a documented transmissible viral load in NYC at the end 2013 included in the validation dataset, the proportions were similar to those in the development dataset (Table 1).

Table 1.

Characteristics and genetic linkage in the development and validation datasets

	Development dataset				Validation dataset
			Linked to ≥1 new infection*				Linked to ≥1 new infection*
	N	Col %	n	Row %	N	Col %	n	Row %
Total	10,609	100.0	538	5.1	8,257	100.0	384	4.7
Sex
Male	7,066	66.6	480	6.8	5,485	66.4	353	6.4
Female	3,543	33.4	58	1.6	2,772	33.6	31	1.1
Race/ethnicity
Black	5,540	52.2	218	3.9	4,555	55.2	165	3.6
Hispanic	3,803	35.8	199	5.2	2,800	33.9	133	4.8
White	1,099	10.4	107	9.7	746	9.0	61	8.2
API	118	1.1	11	9.3	106	1.3	19	17.9
Other	49	0.5	3	6.1	50	0.6	6	12.0
Age
13–24	931	8.8	150	16.1	802	9.7	104	13.0
25–44	4,628	43.6	329	7.1	3,378	40.9	234	6.9
45–64	4,850	45.7	59	1.2	3,827	46.3	46	1.2
65+	200	1.9	0	0.0	250	3.0	0	0.0
Transmission risk
MSM	3,228	30.4	389	12.1	2,715	32.9	296	10.9
IDU	2,010	18.9	17	0.8	1,283	15.5	9	0.7
MSM-IDU	454	4.3	15	3.3	298	3.6	10	3.4
Heterosexual	2,541	24.0	55	2.2	2,039	24.7	35	1.7
Perinatal	398	3.8	5	1.3	409	5.0	4	1.0
Unknown	1,978	18.6	57	2.9	1,513	18.3	30	2.0
Year of diagnosis
Pre-1991	1,279	12.1	6	0.5	838	10.1	3	0.4
1991–1995	1,501	14.1	12	0.8	974	11.8	7	0.7
1996–2000	2,640	24.9	24	0.9	1,856	22.5	17	0.9
2001–2005	2,400	22.6	67	2.8	1,761	21.3	30	1.7
2006–2010	2,789	26.3	429	15.4	1,646	19.9	114	6.9
2011–2013	—	—	—	—	1,182	14.3	213	18.0
Nadir CD4 count in the last 3 years ( cells/mm3)
0–199	5,184	48.9	90	1.7	4,062	49.2	87	2.1
200–349	2,645	24.9	153	5.8	1,906	23.1	105	5.5
350–499	1,702	16.0	160	9.4	1,263	15.3	81	6.4
500+	1,049	9.9	132	12.6	987	12.0	106	10.7
Unknown	29	0.3	3	10.3	39	0.5	5	12.8
Last viral load (copies/mL)
1,500–9,999	3,528	33.3	161	4.6	2,422	29.3	81	3.3
10,000–99,999	4,939	46.6	273	5.5	4,125	50.0	206	5.0
100,000–999,999	2,016	19.0	100	5.0	1,585	19.2	85	5.4
1,000,000+	126	1.2	4	3.2	125	1.5	12	9.6

API, Asian/Pacific Islander; HIV, human immunodeficiency virus; IDU, injection drug users; MSM, men who have sex with men.

^*A new HIV infection was defined as a person diagnosed with HIV at stage 0 or 1 in New York City in the next three calendar years (2011–2013 in the development dataset and 2014–2016 in the validation dataset). The stage of HIV infection was determined based on the revised Centers for Disease Control and Prevention (CDC) classification system.

Among the six selection strategies, the predictive model has the highest AUC (0.86, 95% confidence interval [CI]: 0.84, 0.88), followed by the “Young MSM” (0.79, 95% CI: 0.77, 0.82), “MSM with high viral load” (0.74, 95% CI: 0.72, 0.76), “Random sample of MSM” (0.73, 95% CI: 0.71, 0.76), “Persons with high viral load” (0.56, 95% CI: 0.54, 0.59), and “Random sample” (0.50, 95% CI: 0.48, 0.53) strategies (Figure 1).

Figure 1. Receiver operating characteristic curves and the area under the receiver operating characteristic curve (AUC) for the six selection strategies.

MSM, men who have sex with men; VL, viral load.

Assuming that 500 PLWH with a transmissible viral load could have been selected for a transmission-reduction intervention, there would be striking differences in the characteristics of these PLWH selected by each selection strategy (Table 2). For example, the “Random sample” and “Persons with high viral loads” strategies selected 175 (35.0%) and 158 (31.6%) women, respectively, and the predictive model selected only 10 (2.0%) women. By definition, no women were selected by the three strategies that focused only on MSM.

Table 2.

Characteristics of persons with a transmissible HIV viral load (>1,500 copies/mL) in New York City at the end of 2013 who could have bee selected for transmission-reduction interventions, by selection strategy

	Random sample	Persons with high VL	Random sample of MSM	MSM with high VL	Young MSM	Predictive model
Total	500	500	500	500	500	500
Sex
Men	325	342	500	500	500	490
Women	175	158	0	0	0	10
Race/ethnicity
Black	286	268	235	212	257	196
Hispanic	165	168	163	170	191	200
White	40	54	87	103	31	73
API	5	8	13	10	13	19
Other	4	2	2	5	8	12
Age
13–24	47	49	63	54	347	244
25–44	205	225	274	284	153	255
45–64	228	216	157	154	0	1
65+	20	10	6	8	0	0
Transmission risk
MSM	159	174	447	451	480	415
IDU	86	85	0	0	0	5
MSM-IDU	16	13	53	49	20	15
Heterosexual	117	122	0	0	0	29
Perinatal	27	21	0	0	0	0
Unknown	95	85	0	0	0	36
Year of diagnosis
Pre-1991	52	43	42	27	0	0
1991–1995	59	53	27	32	0	0
1996–2000	114	109	78	86	0	0
2001–2005	118	103	99	100	14	0
2006–2010	92	96	132	125	194	9
2011–2013	65	96	122	130	292	491
Last viral load in 2013 (copies/mL)
1,500–9,999	165	0	143	0	101	96
10,000–99,999	245	0	248	0	288	296
100,000–999,999	86	375	98	451	103	92
1,000,000+	4	125	11	49	8	16
Genetically linked to ≥1 new HIV infection in 2014–2016*
No	483	471	462	440	384	363
Yes	17	29	38	60	116	137
Genetic linkage rate (%)†	17/500 = 3.4%	29/500 = 5.8%	38/500 = 7.6%	60/500 = 12.0%	116/500 = 23.2%	137/500 = 27.4%
Rate ratio	1.00	1.71	2.24	3.53	6.82	8.06
95% CI	—	0.95, 3.06	1.28, 3.91	2.09, 5.96	4.17, 11.18	4.95, 13.13

API, Asian/Pacific Islander; CI, confidence interval; HIV, human immunodeficiency virus; IDU, injection drug users; MSM, men who have sex with men; VL, viral load.

^*A new HIV infection was defined as a person diagnosed with HIV at stage 0 or 1 in New York City in 2014–2016. The stage of HIV infection was determined based on the revised Centers for Disease Control and Prevention (CDC) classification system.

^†Genetic linkage rate was defined as the proportion of PLWH with a transmittable viral load (>1,500 copies/mL) at the end of 2013 whose viral sequence was genetically linked to that of at least one new HIV infection in 2014–2016.

In terms of age, the proportion of PLWH 45 years or older selected for intervention was 49.6% by the “Random sample,” 45.2% by the “Persons with high viral loads,” 32.6% by the “Random sample of MSM,” 32.4% by the “MSM with high viral loads,” 0% by the “Young MSM” strategy, and 0.2% by the predictive model.

Among 500 PLWH at the end of 2013 selected by the “Random sample” strategy, 17 of them were linked to at least one new HIV infection diagnosed in NYC in 2014–2016, with a genetic linkage rate of 3.4% (17/500). The genetic linkage rates were 5.8%, 7.6%, 12.0%, 23.2%, and 27.4%, respectively, for the “Persons with high viral loads,” “Random sample of MSM,” “MSM with high viral loads,” and “Young MSM” strategies, and the predictive model. The predictive model had the highest genetic linkage rate and was 8.06 (95% CI: 4.95, 13.13) times higher than the “Random sample” strategy.

Figure 2 shows the genetic linkage rate of each strategy by number of PLWH selected for transmission-reduction interventions. When the number of PLWH selected for interventions increased from 250 to 1,000, no changes in the genetic linkage rate were observed for the four strategies with the lowest rates, i.e., the “Random sample,” “Persons with high viral loads,” “Random sample of MSM,” and “MSM with high viral loads” strategies. The genetic linkage rate for the top two strategies decreased, and the gap between them narrowed, as the number of PLWH selected for intervention increased. When 250 PLWH were selected for transmission-reduction interventions, the genetic linkage rates were 32.4% and 27.6%, respectively, for the predictive model and the “Young MSM” strategy, and the gap between them was 4.8 percentage points (95% CI: −3.5, 13.0); when 1,000 PLWH were selected, the genetic linkage rates were 22.2% and 20.0%, respectively, and the gap between them was 2.2 percentage points (95% CI: −1.5, 5.8), with a difference-in-difference of 2.6 percentage points (95% CI: −5.6, 10.8).

Figure 2. Genetic linkage rate, by selection strategy and number of PLWH with a transmissible viral load (>1,500 copies/mL) selected for transmission-reduction interventions*^†.

HIV, human immunodeficiency virus; MSM, men who have sex with men; PLWH, persons living with human immunodeficiency virus; VL, viral load.

*Genetic linkage rate was defined as the proportion of PLWH with a transmissible viral load (>1,500 copies/mL) at the end of 2013 whose viral sequence was genetically linked to that of at least one new HIV infection in 2014–2016.

†A new HIV infection was defined as a person diagnosed with HIV at stage 0 or 1 in New York City in 2014–2016. The stage of HIV infection was determined based on the revised Centers for Disease Control and Prevention (CDC) classification system.

As expected, the sensitivity analysis results showed that the genetic linkage rate decreased as the genetic distance threshold decreased from 0.015 to 0.005 substitution/site and genetic linkages were limited to acute HIV infections (Table 3). However, the rate ratio stayed relatively stable, and the predictive model always had the highest ratio.

Table 3.

Sensitivity analysis results—number of PLWH with a transmissible viral load (>1,500 copies/mL) whose viral sequence was genetically linked to that of at least one new HIV infection or acute HIV infection in the next three calendar years, by selection strategy, genetic distance, and stage of new HIV infection^*†

	Random sample	Persons with high VL	Random sample of MSM	MSM with high VL	Young MSM	Predictive model
Total number of persons selected for a transmission-reduction intervention (N)	500	500	500	500	500	500
Genetically linked to ≥1 new HIV infection within a 1.5% genetic distance*
n	17	29	38	60	116	137
Genetic linkage rate (%)§	17/500 = 3.4%	29/500 = 5.8%	38/500 = 7.6%	60/500 = 12.0%	116/500 = 23.2%	137/500 = 27.4%
Rate ratio	1.00	1.71	2.24	3.53	6.82	8.06
95% CI	—	0.95, 3.06	1.28, 3.91	2.09, 5.96	4.17, 11.18	4.95, 13.13
Genetically linked to ≥1 new HIV infection within a 0.5% genetic distance*
n	7	9	15	22	31	57
Genetic linkage rate (%)§	7/500 = 1.4%	9/500 = 1.8%	15/500 = 3.0%	22/500 = 4.4%	31/500 = 6.2%	57/500 = 11.4%
Rate ratio	1.00	1.29	2.14	3.14	4.43	8.14
95% CI	—	0.48, 3.43	0.88, 5.21	1.36, 7.29	1.97, 9.96	3.75, 17.68
Genetically linked to ≥1 acute HIV infection within a 1.5% genetic distance†
n	6	15	17	31	43	58
Genetic linkage rate (%)§	6/500 = 1.2%	15/500 = 3.0%	17/500 = 3.4%	31/500 = 6.2%	43/500 = 8.6%	58/500 = 11.6%
Rate ratio	1.00	2.50	2.83	5.17	7.17	9.67
95% CI	—	0.98, 6.39	1.13, 7.13	2.18, 12.27	3.08, 16.68	4.21, 22.20
Genetically linked to ≥1 acute HIV infection within a 0.5% genetic distance†
n	2	5	7	12	12	17
Genetic linkage rate (%)§	2/500 = 0.4%	15/500 = 1.0%	7/500 = 1.4%	12/500 = 2.4%	12/500 = 2.4%	17/500 = 3.4%
Rate ratio	1.00	2.50	3.50	6.00	6.00	8.50
95% CI	—	0.49, 12.82	0.73, 16.76	1.35, 26.67	1.35, 26.67	1.97, 36.59

CI, confidence interval; HIV, human immunodeficiency virus; PLWH, persons living with HIV; VL, viral load.

^*A new HIV infection was defined as a person diagnosed with HIV at stage 0 or 1 in New York City in 2014–2016. The stage of HIV infection was determined based on the revised Centers for Disease Control and Prevention (CDC) classification system.

^†An acute HIV infection was defined as a person diagnosed with stage 0 HIV infection in New York City in 2014–2016. The stage of HIV infection was determined based on the revised CDC classification system.

^§Genetic linkage rate was defined as the proportion of PLWH with a transmissible viral load (>1,500 copies/mL) at the end of 2013 whose viral sequence was genetically linked to that of at least one new HIV infection or acute HIV infection diagnosed in 2014–2016.

DISCUSSION

The CDC recommends that jurisdictions include both HIV-negative persons and PLWH in interventions to reduce HIV transmission.^[2] Since PLWH with an undetectable viral load cannot transmit HIV to their sexual partners,^{[6, 28]} it would be more effective to focus on PLWH with a transmissible viral load. The majority of PLWH in the United States receive regular HIV care, including viral load monitoring.^{[29, 30]} Therefore, it is not difficult to identify PLWH with a transmissible viral load in state and local surveillance systems. The challenge is how to prioritize them effectively. Using HIV sequence data, we developed a predictive model to select PLWH with a transmissible viral load for transmission-reduction interventions and have shown that the model performs better than all other selection strategies included in our analysis. The predictive model should perform even better when the program has more limited resources and must select fewer PLWH for transmission-reduction interventions.

Besides its better performance, the predictive model has a number of advantages. First, it can systematically select PLWH for interventions by considering multiple factors simultaneously, while other selection strategies focus on only one or two factors and may not prioritize individuals at the greatest risk for transmission.

Second, unlike other selection strategies, the predictive model does not limit PLWH with certain characteristics for intervention. For example, women and heterosexual men would not be selected by a selection strategy that focuses only on MSM, but they could be selected by the predictive model if other factors put them at a higher risk of transmitting HIV, such as young age, high viral loads, or history of drug use.^{[31, 32]}

Third, HIV sequence data are used to build the predictive model, but after the predictive model is built, HIV sequence data are not needed for the model to identify PLWH for transmission-reduction interventions, although additional sequence data can be used to refine the model. At the end of 2013, there were 10,128 PLWH with a transmissible viral load in NYC, of whom 8,257 (81.5%) had HIV sequence data and 1871 (18.5%) did not. Applying the predictive model to select 500 PLWH for intervention, 375 (75.0%) would be selected from those with sequence data and 125 (25.0%) from those without. The higher probability of being selected from those without sequence data (125/1,871 = 6.7% vs. 375/8,257 = 4.5%) suggests that some factors, e.g., gender, age, race/ethnicity, and CD4 count at diagnosis,^[33] that make PLWH less likely to be genotyped may also put them at a higher risk of transmitting HIV.

The predictive model also has limitations. First, our definition of new infection is based on the CD4 count at the time of diagnosis and there could be some misclassifications—new infections with a CD4 count at diagnosis lower than 500 copies/mL (false negatives) and established infections with a CD4 count at diagnosis higher than 500 copies/mL (false positives). To predict a new transmission, minimizing false positives is more important than minimizing false negatives. We were able to further minimize false positives by conducting a sensitivity analysis among people with an acute HIV infection and reached the same conclusion—the predictive model performed better than the other five selection strategies.

Second, we used genetic links to evaluate each selection strategy, and genetic links alone, particularly at this liberal distance threshold, cannot be used to represent direct transmissions. Although we are unable to confirm direct transmissions at the individual level, it is reasonable to conclude at the population level that a subpopulation with more genetic links represents more transmissions. Selecting such a subpopulation for intervention would be an effective way to prevent onward transmission.

Third, to keep the model simple, PLWH whose viruses were genetically linked to that of at least one new infection in the next three calendar years were all treated equally in the model building, despite the fact that some were linked to more than one new infection. Building the model this way should not affect our conclusion that the predictive model performs better than the other five selection strategies, because, 1) it was common (41.7%) for PLWH to link to more than one new infection, 2) it was also common (28.7%) for new infections to link to more than one PLWH, and 3) the same method was applied to all six selection strategies.

Fourth, HIV-TRACE requires a minimum of 500 nucleotides to calculate the genetic distance between two sequences. The sequences included in our analysis have a length between 669 and 1,600 nucleotides, with a median length of 1,212 (interquartile range [IQR]: 1,212, 1,497). Different lengths of sequences may have an impact on the distance calculation, but they should have little impact on our conclusion, because the same genetic distance calculation method was applied to all six selection strategies.

Fifth, the current predictive model can only be used to prioritize viremic PLWH who are in care, i.e., it cannot be used to prioritize out-of-care patients for transmission-reduction interventions because their viral load data are not available. However, since out-of-care patients are eligible for re-engagement services by health departments that do “Data to Care” work, they have other opportunities to be selected for HIV interventions.^[34]

Finally, the analysis was conducted in NYC, where HIV surveillance data may be more complete than other jurisdictions, including the high proportion of diagnosed infections, the high percentages of diagnosed PLWH entered in the registry, PLWH with sequence data, and PLWH with complete information on the variables included in the model. It is also possible that PLWH in NYC have different relative transmission rates by the categories we examined than those elsewhere. The purpose of this analysis is to demonstrate that predictive modeling using case and molecular surveillance data can be used to prioritize persons with a transmissible HIV viral load for transmission-reduction interventions. Jurisdictions should evaluate the quality and completeness of their data before using our method to build their own model for transmission-reduction interventions.

Using case and molecular surveillance data, we developed a predictive model to prioritize PLWH with a transmissible viral load for interventions to reduce onward transmission and found the model to perform better than other selection strategies. We suggest investigating applying this method in the real world to evaluate its feasibility and effectiveness. Before implementation, jurisdictions also need to consider the ethical implications of selecting persons with specific, readily identifiable characteristics, e.g., race/ethnicity, sexual orientation, and transmission risk, for targeted interventions, and evaluate its latent consequences, possibly in consultation with a community advisory board that is sensitive to community concerns about stigma and discrimination against PLWH. Both external and internalized stigma may drive people away from the services that they need.