This study was based on a well-run, long-term ART programme involving a wide range of public hospitals throughout Thailand, with good quality data collection including accurate ascertainment of AIDS diagnoses and cause of death for most patients. Consistent with other studies from low and high-income settings, CD4 count was found to be the strongest predictor of death and of new ADE 
. In our cohort, the decline in mortality rate over time from cART initiation (well-documented particularly in low-income settings 
) was accounted for by current CD4 count, further underlining its key role as an immediate prognostic indicator. As expected, mortality rate at CD4 count <100 cells/mm3
was substantially high (124.8 per 100 person-years). Therefore, patients with low CD4 count, even if virologically suppressed, should be closely monitored, with underlying causes of immunosuppression (including potentially non-HIV related conditions) promptly investigated and addressed.
We observed improved survival with increasing CD4 count even at levels above 200 cells/mm3
, with mortality being rare at ≥500 cells/mm3
(rate 1.1 per 1000 person-years). A recent analysis of four large treatment cohorts from Sub-Saharan Africa reported a 1.7 fold increase in mortality risk at CD4 levels 350–499 cells/mm3
compared to ≥500 cells/mm3
(adjusted for only baseline characteristics and calendar period), comparable to our estimate of 2.5 (95% CI 1.9–3.4) 
. Furthermore, a French study showed patients on long-term cART with CD4 count >500 cells/mm3
had similar mortality rates as the general population 
. These findings provide growing evidence of the benefit of maintaining CD4 count above 500 cells/mm3
while on treatment. They also suggest that relaxing monitoring for patients on treatment with high stable CD4 count should be evaluated for settings with limited capacity for CD4 monitoring. In addition, early HIV diagnosis and early treatment initiation before onset of severe immunosuppression are indicated, given that high CD4 count at ART initiation is associated with improved long-term CD4 cell count recovery 
; the effect on long-term outcomes of starting ART at CD4 count >500 cells/mm3
compared to deferring until CD4 falls below 350 cells/mm3
is being investigated in the ongoing randomized START trial 
Previous studies across different settings already showed the independent association between low haemoglobin on cART and disease progression 
, as also observed in our study. The link between low haemoglobin and mortality is not well-understood and appears multifactorial, but is likely to be partly due to anaemia being an indicator of other co-morbidities, particularly tuberculosis 
. Consistent with this, we found 80% of patients dying of tuberculosis had haemoglobin level <10 g/dl at the time of death, compared with 35% of deaths due to other infections and none of those due to non-infection related causes. Although haemoglobin level was lower in women throughout follow-up in our cohort, its prognostic value did not vary by gender (data not shown), consistent with results from a previous UK study 
In our cohort, high current viral load was weakly associated with both mortality and new ADE but only at ≥100,000 copies/ml, consistent with a previous ART-CC study 
. Other studies from low- and high-income settings generally reported a modest association between lack of viral load suppression and poor prognosis 
. In a large cohort from South Africa, viral load was more strongly prognostic after 12 months, suggesting that viral load monitoring may be more informative after 1 to 2 years of cART 
. Given the need to prevent accumulation of resistance on a virologically failing regimen, development of viral assays which are cheap, practical and reliable for use in low-income settings remains important for improving patient monitoring 
. Of interest, high viral load at cART initiation remained predictive of new ADE/death (but not of death) after adjusting for subsequent viral load and CD4 measurements. The explanation for this association, which has not been previously reported, is unclear but could partly be due to delay in diagnosis of certain pre-existing conditions at the time of cART initiation.
Substantial variation in the impact of different ADEs on mortality was previously observed in an ART-CC analysis including treated patients in Europe and US 
. When applying the prognostic categories of ADEs proposed from this study to our cohort, there was in fact little difference in the effect of mild compared with moderate/severe ADEs on mortality. Our analysis however only included one patient with a severe ADE (progressive multifocal leukoencephalopathy) after starting cART. Furthermore, due to sparse data, we did not evaluate effects of individual ADEs (as was done in the ART-CC analysis), but considered the first event occurring within each ADE prognostic category.
In our study, rate of loss-to follow-up was relatively low, with efforts made to trace patients missing clinic visits using telephone calls and home visits. Although voluntary patient withdrawal from the study was usually due to relocation or ART being accessed elsewhere through the national treatment programs, it is important to note that patients who were either lost to follow-up or voluntarily withdrew were less likely to be virologically suppressed after cART initiation compared to those still in follow-up, with loss to follow-up also associated with lower CD4 attained (data not shown). This could lead to under-estimation of the effects of CD4 count and viral load if patient drop-out was associated with mortality or progression to new ADE’s, conditional on the covariates in the models (including current CD4 count and viral load).
A limitation to our analysis was that effects of covariates were not assessed separately for AIDS-related and non AIDS-related mortality due to limited number of deaths. An association between CD4 count and non-AIDS related mortality has been reported, albeit of weaker magnitude compared with that for AIDS-related mortality 
. This is consistent with our observation that patients dying of causes not related to infection had higher CD4 count and haemoglobin at the time of death compared to those dying of infection(s), and were also more likely to be virologically suppressed. Of note, we found patients were more likely to die of non-infection related than infection-related causes after 2 years of cART, indicating that ART monitoring potentially needs to account for the changing patterns of mortality over time. Another issue is that most patients with early progression to death and to new ADE were excluded from analyses owing to lack of laboratory measurements after starting cART; risk factors of early mortality in our cohort have been previously described 
. Consequently, only 14% of new ADEs and 8% of deaths occurred within 6 months of cART initiation, and while immune reconstitution syndrome (IRIS) may account for some of these early events, it is unlikely to play a significant role in our overall findings.
In conclusion, immunosuppression and anaemia may reflect the presence or development of conditions that lead to mortality. Half of the deaths in this Thai cohort were attributable to infections which were chronic and generally considered treatable, while the cause of death remained unknown in nearly a fifth despite regular follow-up. This suggests a significant proportion of patients died of conditions which were not diagnosed in time or appropriately treated, possibly due to constraints in this resource-limited setting and lack of experience in handling such complications at the beginning of the program when the priority was given to starting antiretroviral treatment for all those who urgently needed it. Patients with low CD4 count or haemoglobin should therefore receive more intensive case-management, while monitoring could potentially be relaxed for those with high CD4 if resources are limited. ART monitoring in low-income settings remains a research priority, and in particular, better understanding of predictors of morbidity and mortality due to non-AIDS disease is needed given patients are surviving longer on treatment.