Pablo Lamelas –
Hypertension is the largest responsible of death in the world. Its control and risk reduction is a health priority, specially middle- and low-income countries where risk factor control and mortality is worse compared to higher income countries. HOPE-4 was a very complex and ambitious initiative, that started assessing the barriers for hypertension control and risk modification in Colombia an Malaysia. Then, an intervention package was created based on the detected barriers and adapted to local context. The nature of the intervention and setting made this study particularly challenging, and feel proud of what the team has achieved, also a unique learning opportunity regarding research methods.
HOPE-4 trial summary
In 1,371 patients with uncontrolled blood pressure (high blood pressure, with or without prior hypertension diagnosis) belonging to urban and rural communities from Colombia and Malaysia, communities were randomized to receive a multifaceted intervention package or usual care. The intervention included: 1) non-physician health workers (NPHW) to perform physician tasks (assessment and management of risk factor control) at the community setting (outside hospitals or clinics), 2) free anti hypertensive medication dispensation and recommendation by NPHW supervised by physicians, and 3) family/friend treatment supporter. The primary outcome was risk modification based on Framingham risk score at 12 months. The study demonstrated a drastic fall in Framingham score in the intervention group, as well as other important secondary outcomes like blood pressure, cholesterol, medication adherence, and other lifestyle behaviours. This intervention could make a huge impact in developing countries, and likely scalable to lower income areas in high income countries.
This trial randomized communities instead of individuals. This randomization method was selected for two main reasons. The most important was to avoid contamination of the intervention in control patients. If individuals from the same community were allocated to intervention or usual care, since the intervention included free dispensation of medications and lifestyle changes, control participants may get “contaminated” with the intervention delivered in those allocated to intervention.
The second reason was logistics. The NPHWs screened all the communities, but then they made frequent follow-up in intervention communities only. Therefore, logistically speaking, less NPHWs and less transfers would be needed to properly deliver the intervention at the community level.
What are the downsides of this cluster randomization design? Since the unit of randomization is the community (a cluster) while the unit of analysis are individuals, the individuals from each community tend to correlate with other individuals from the same community. In other words, any variable you may imagine from people belonging to community A is more likely to be better correlated (a little or a lot) with other individuals from community A rather than the rest of the communities (B, C, D, E, etc).
The statistical way to calculate this correlation is the Intraclass Correlation Coefficient (better known as ICC) which represents the percent of the total variance observed explained by between cluster differences. For example: blood pressure ICC is usually close to 0.05 (or 5%), meaning that 5% of the total variance observed (between and within clusters) is explained by between cluster differences. The higher the ICC, higher correlation between the variable of the individual with the cluster they belong to. The higher the ICC, the smaller the effective sample size of the randomized trials, meaning that you have to enrol larger number of patients to preserve the power of an individual patient randomization scheme.
How we dealt with this issue? We had data from similar urban and rural communities from the Prospective Urban Rural Epidemiology study (PURE), so we were able to calculate a proxy of the expected ICC for each variable of interest, like Framingham, blood pressure and cholesterol levels, making our sample size calculations more predictable in terms of ICC.
The study randomized communities after screening. This lead to some drop outs or even deaths from the time patients were screened until the community was randomized. Why randomization was not done before?
The main reason was to avoid selection bias: if communities were randomized before they were screened, the patients will know the allocation and participate or not in the study according to allocation. Therefore, screen and then randomization was preferred to avoid selection bias.
Large drop in risk among control communities
You noticed that there was quite a drop in Framingham risk among control communities, even though they were assigned to “usual care”. In normal circumstances, in the “real-world”, Framingham risk goes up as time goes by, instead of down. Until more explanations come along, I have at least two potential explanations that future analyses may dig better in understanding this.
First, some degree of contamination during the screening phase. The NPHWs performed screening in the same way in control or intervention communities, since screening was done before randomization. This study included a considerable amount of patients that did not know they had hypertension. Then, many patients were diagnosed with hypertension (23%) in the usual care group went to their doctor to initiate treatment.
Also, if patients had severely elevated blood pressured, as per IRB recommendation, participants were suggested to attend the emergency room to check their blood pressure and likely start therapy. Then, although the screening phase before randomization increased the validity of the results in terms of selection bias, it also caused that some participants in the usual care group improved risk factors control during the screening phase.
Second, regression to the mean may also play a role. If the mean blood pressure of the screened population is <140 mmHg and >140 mmHg was used to consider high blood pressure as study inclusion criteria, the real blood pressure go, in average, towards the mean in repeated measures during follow-up. In other words, some people enrolled in the trial may have a “real” blood pressure <140 mmHg.
This is called regression to the mean, and usually affect analysis in which the mean is under the threshold selected to be include in the study. This is one the reasons to have a control group: if doing a single cohort study, pre and post intervention, you may see a drop in blood pressure among people with high blood pressure >140 mmHg and just doing nothing rather than doing a repeated visits in the next months with regression to the mean effect.
Important point: Although the proper way to scientifically prove the intervention effect is through a controlled randomized trial, the effect of this intervention (if scaled to real world) likely to better be represented by pre- post- estimates rather than between group (intervention vs control) estimates. Why? Mainly because the large drop in risk seen in the control group likely explained by some degree of contamination and regression to the mean, therefore bias, diluting the real effect of the intervention. This effect dilution was a price to pay to make results more valid by avoiding selection bias if screening was done after randomization (this is a personal opinion).
Surrogate outcomes in an open label study
Surrogate outcomes has some limitations when interpreting open label studies. The participants nor the NPHWs could be blinded to the intervention given the nature of the trial. Therefore, is impossible to rule out other co-interventions not planned in the trial, as well as ascertainment bias by those collecting the data. Blood pressure was measured using automated devices and taken using protocols, and blood lipids were analyzed in core laboratories, making bias less likely to happen. Furthermore, mean fasting time for blood analysis was the same between intervention and control communities, suggesting no significant performance bias.
Both blood pressure reduction in hypertensive patients and blood cholesterol in people at risk of cardiovascular events are well stablished surrogate outcomes: are know to have a causal association with clinical outcomes and their modification with treatment correlates with lower adverse events. Still are not patient-important outcomes, which will need much large sample size and longer follow-up.
Why Framingham score as primary outcome instead of blood pressure or cholesterol? The multifaceted intervention of HOPE-4 goes beyond blood pressure control and cholesterol reduction, it also impacts smoking status. Then, the best way to combine these three aspect in one estimate was by using Framingham score.
Hypertension is the leading risk factor for premature death, which remains uncontrolled in most regions of the world. The HOPE-4 initiative developed an evidence-based multifaceted package to improve risk factor modification in this population. This trial confirmed the implementability as well as the efficacy of reducing cardiovascular risk in vulnerable regions of the world. This intervention has the power to substantially reduce cardiovascular risk by scales never seen before.
Sometimes cardiologists get excited/interested in explanatory trials of specific diseases with modest reduction in events, losing perspective of the real impact on the cardiovascular burden we see today. HOPE-4, although not a study with several thousand participants followed for many years and showing death reduction, still has a great potential for controlling cardiovascular disease, specially in middle- and low- income countries.
Do not want to miss the chance to thank the HOPE-4 team, specially JD Schwalm who has been a great mentor during my stay at PHRI along with Dr Yusuf. For those who do not know JD, he is also a plumber with interests in global health: plumbers want less coronary stenosis!