Prognostic indicators in adults hospitalized with falciparum malaria in Western Thailand

Background Severe malaria remains a major cause of death and morbidity amongst adults in the Asiatic tropics. Methods A retrospective analysis of the clinical and laboratory data of 988 adult patients, hospitalized with Plasmodium falciparum malaria and prospectively recruited to malaria studies in western Thailand between 1986 and 2002, was performed to assess the factors associated with a fatal outcome. Different severity scores and classifications for defining severe malaria were compared and, using multiple logistic regression, simple models for predicting mortality developed. Results The proportion of patients fulfilling the WHO 2000 definition of severe malaria was 78.1%, and their mortality was 10%. Mortality in patients given parenteral artesunate or artemether (16/317, 5%) was lower than in those given parenteral quinine (59/442, 13%) (P < 0.001). Models using parameter sets based on WHO 1990, 2000 and Adapted AQ criteria plus blood smear parasite-stage assessment gave the best mortality prediction. A malaria prognostic index (MPI), derived from the dataset using five clinical or laboratory variables gave similar prognostic accuracy. Conclusions The mortality of severe malaria in adults has fallen and the switch from quinine to artesunate has probably been an important contributor. Prognostic indices based on WHO 2000 definitions, and other simpler indices based on fewer variables, provide clinically useful predictions of outcome in Asian adults with severe malaria.


Background
Plasmodium falciparum malaria still kills~0.7 million people each year. The majority of those who die are children in sub-Saharan Africa, but~43,000 patients still die each year in the Asia-Pacific [1]. Admission clinical features that predict death have been defined in African children [2][3][4][5][6]. For Asian adults the relationships between mortality and a wide range of individual variables have been tested, including parasitaemia and parasite-stage distribution and intraleucocytic malaria pigment on blood smears, renal failure, hypoglycaemia, cerebral malaria, acidosis and older age [7][8][9][10][11][12][13][14] (Additional file 1, Tables 1 and 2). However, there have been few assessments of which combinations of prognostic factors are the best predictors of mortality. Although a single definition may not be appropriate everywhere, it is important to attempt to define severe malaria to aid clinicians in recognizing those at risk of death and in need of parenteral therapy. This will inform decisions on whether transfer to a higher level of care, if available, is appropriate, facilitate comparison between datasets, longitudinal epidemiological surveillance, and robust case definitions in evaluation of interventions [11,15].
Several definitions, classifications and severity scores have been proposed. The World Health Organization (WHO) has produced three guidelines which include definitions of severe malaria (Additional file 1, Tables 1 and 2) [16][17][18]. A slightly stricter definition of severe malaria (' AQ') was developed for a clinical trial in Vietnam [19] and adapted (with the addition of plasma lactate and serum bicarbonate measurements) for a trial in Thailand Respiratory distress, requiring mechanical ventilation BCAM score uses serum bicarbonate as a marker of acidosis with cut-off values of ≥24 mmol/L (score = 0) for normal, ≥15-< 24 mmol/L (score = 1) for deranged, and <15 mmol/L (score = 2) for very deranged along with thresholds for coma of GCS ≤14 and GCS ≤10 [23]. The acidosis and coma scores were summed to give the BCAM score, ranging from 0-4. RCAM score uses respiratory rate as a surrogate marker for acidosis with cut offs of <20 breaths/min (score = 0) for normal, 20-39 breaths/min (score = 1) for deranged, and ≥40 breaths/min (score = 2) for very deranged.
[20] and the subsequent multicentre SEAQUAMAT trial which enrolled 1461 Asian patients [21]. The malaria severity assessment score (MSA), based on haemoglobin, serum creatinine, requirement for mechanical ventilation and Glasgow coma score (GCS) was developed in central India [22]. Recently, in order to simplify and, therefore, broaden usage, the coma acidosis malaria (CAM) score was developed [23], based on data from Asian adults with severe malaria [21]. A score of <2/5, when tested with data from different studies in Vietnamese and Bangladeshi adults, had a positive predictive value (PPV) for survival of 94-95%, suggesting that these patients could be cared for without admission to an intensive care unit (ICU).
The clinical and laboratory features of adults with falciparum malaria recruited prospectively to hospitalbased studies on the western border of Thailand between 1986-2002, conducted by the Mahidol University-Oxford Tropical Medicine Research Unit and colleagues, were analysed. The specificity and sensitivity of the different severe malaria definitions and scoring systems were examined and simple models to identify adults at risk of death built.

Study sites, years and studies
Data from 988 adult (≥15 years) patients with asexual forms of P. falciparum present on peripheral blood slides, admitted to hospital with malaria and then recruited to clinical research studies on the western border of Thailand between 1986 and 2002 were analysed. Patients were recruited at hospitals in Kanchanaburi (1986-1993; n = 571), Sangklaburi (1994-1995; n = 74), and Mae Sot (1995-2002; n = 343) and described in a series of papers [24]. All patients gave informed consent to participation and all studies were approved by the Ethics Committee of the Faculty of Tropical Medicine, Mahidol University and/or the Ethical and Scientific Review Subcommittee of the Ministry of Public Health, Government of Thailand.

Clinical and laboratory assessment
As patients were recruited to a variety of studies, clinical and laboratory evaluations varied. All patients had a full history and examination performed and haematocrit and parasitaemia determined. Thick and thin blood films were stained immediately with Field's stain and parasites counted and staged [7,8]. Admission blood samples for full blood count, serum sodium, potassium, creatinine, urea, total bilirubin, direct bilirubin, alkaline phosphatase, alanine transaminase (ALT), aspartate transaminase (AST), plasma lactate, glucose and, except in 1994, for plasma bicarbonate. Lumbar punctures were performed for the majority of patients with reduced GCS and cerebrospinal fluid cell, protein and glucose concentrations determined and Gram stains examined.

Management
The anti-malarial treatment regimens used in clinical studies in western Thailand changed over the 18 years as new anti-malarials were introduced and were: 1. Non-artemisinin-based parenteral treatment: Intravenous quinine dihydrochloride with or without a 20 mg salt/kg or 7 mg/kg loading dose followed by 10 mg/kg every 8 h followed by oral quinine salt 10 mg/kg every 8 h, combined, when the patient was able to take oral medication, with oral quinine alone or with oral tetracycline, doxycycline, mefloquine (alone or combined with sulphadoxinepyrimethamine (SP)) or single dose primaquine to give a total treatment course of 7 days. 2. Non-artemisinin-based oral treatment: Oral quinine sulphate 10 mg salt/kg every 8 h alone or with oral tetracycline, doxycycline, mefloquine (alone or combined with SP), single dose primaquine or proguanil to give a total treatment course of 7 days. Or oral mefloquine (alone or combined with SP) at 15 mg base/kg on the first day with or without 10 mg/kg on the next day Or oral halofantrine (8 mg/kg at 0, 6, 12 h) with tetracycline plus or minus oral quinine 3. Artemisinin-based parenteral treatment: Intravenous artesunate (Guilin Pharmaceutical Factory No. 2, Guangxi, People's Republic of China); 2.4 mg/kg stat, 1.2 mg/kg at 12 h followed by 1.2 mg/kg every 24 h) or intramuscular artemether (3.2 mg/kg stat followed by 1.6 mg/kg every 24 h) for 7 days alone or with oral doxycycline, tetracycline or mefloquine (alone or combined with SP) with or without primaquine. Sixty-nine patients were treated with iv artesunate combined with iv quinine followed by the above oral regime [25]. 4. Artemisinin-based oral treatment: Oral artesunate or artemether (4 mg/kg for 3 days or 2 mg/kg for 7 days) alone or with oral doxycycline, tetracycline or mefloquine (alone or combined with SP) with or without primaquine. Eighteen patients received oral dihydroartemisinin (4 mg/kg) in replacement for one artesunate daily dose [26].
Supportive treatment was in accordance with guidelines [17,27]. Facilities for urinary catheter and nasogastric tube placement, blood transfusion and lumbar puncture were available at Sangklaburi Hospital. In addition, mechanical ventilation, peritoneal dialysis and central venous access were available at Kanchanaburi and Mae Sot.

Statistical analysis and modelling
Six main models have been used to define severe malaria in adults, those published by WHO and the AQ, MSA and CAM scores [16][17][18][19][20][21][22][23] (Additional file 1). The accuracy of the APACHE II score [28] could not be assessed as not all necessary laboratory variables were measured, and the definition of WHO 1986 was not evaluated [16]. Data were not collected specifically for this analysis and in order to examine the sensitivity and specificity of WHO 1990 and 2000 definitions [17,18], minor adaptations were made to allow assessment (Additional file 1, Tables 1 and 2). Where the WHO 2000 definition [18] does not include quantitative cut offs, the WHO 1990 [17] cut offs have been used. As blood pH, and hence base deficit, data were not available, calculation of the CAM score was not possible and so the modified CAM scores, BCAM and RCAM were evaluated, using serum bicarbonate and respiratory rate, respectively, as surrogates of acidosis (Tables 1 and 2).
Statistical analyses were performed using Stata (v11.0; Stata Corporation, USA). All univariate comparisons between survivors and patients who died were performed using logistic regression and adjusted for study site. Ten sets of clinical and laboratory assay variables (Tables 1  and 2) were used to construct diagnostic rules to predict death. Logistic regression analysis with a stepwise forward variable selection procedure was employed to find independent predictors of death at P < 0.055, and P ≥ 0.055 for entry and removal, respectively. Fractional polynomials [29] were used to test for non-linear relationships between outcome variable and continuous covariates. All models were adjusted for artemisinin-based combination therapy (ACT) and study site. The identified model was rerun on the maximum available sample size and each of the variables not in the model were tested for inclusion using the Wald test. The predictive utility of each final model was assessed by receiver operating characteristic (ROC) curve analysis.

Malaria prognostic index
Variables selected into any final logistic regression model, based on published criteria (see above), were used to define the malaria prognostic index (MPI). Since for all logistic models (above), inclusion of study site as an independent variable did not improve the model nor change the co-efficients for other covariates significantly, it was not included in the development of the MPI. Each variable was categorized into four groups using rounded quartiles or commonly used cut offs (as for GCS). Univariate logistic models (with sets of corresponding binary variables) were fitted and categories with similar (P > 0.05) odds ratios (OR) were grouped together. Co-efficients of the final multivariate model were rounded to the nearest integer and used to calculate for each patient a linear combination of variables (i. e, sum of variables multiplied by the rounded co-efficients)the MPI. The MPI was calculated for each patient and the ROC analysis used to evaluate its prognostic utility.
The predictive value of the MPI was further evaluated using cross validation [30] on a subset of data with no missing values for variables chosen to define MPI. Each observation in this subset was sequentially removed, logistic regression with stepwise variable selection was performed on the remaining n-1 observations using all categorized variables, and the final model was used to calculate sensitivity, specificity on n-1 observations and classification results for the one excluded observation. Co-efficients were rounded to integers and cut offs for the linear predictor between two and six were used. For each cut off, classification results for observations excluded from the subsequent models were used to calculate jacknife estimates of sensitivity and specificity.
Mortality did not significantly differ between the three sites (9% in Kanchanaburi, 3% in Sangklaburi and 8% in Mae Sot; P = 0.20). Artemisinin derivatives were given to 7% of patients in Kanchanaburi, 77% in Sangklaburi and 83% in Mae Sot (P < 0.001), reflecting temporal changes in treatment policy and study protocols. Patients in Kanchanaburi had received prior malarial treatment more frequently than at other sites and women there were more often pregnant. Patients in Mae Sot presented with lower coma scores (69% with coma score of 15 compared to >80% at the other sites and 20% with coma score ≤11 compared to 9-10% in the other sites), higher parasitaemia, higher proportions of trophozoites on admission film and a greater likelihood of a palpable liver or spleen (Tables 3, 4, 5, and 6). However, other key variables such as haematocrit, bicarbonate and lactate were similar across all sites.
Forty-one admission variables were significantly associated (P < 0.05) with death on bivariate analysis (Tables 7  and 8). Variables previously associated with mortality that were not significantly associated with death in our cohort were patient age, the number of days of illness before  admission, and plasma glucose. Although there was no apparent overall relationship between mortality and age, mortality was higher with increasing age for those treated with quinine (OR (95% CI) 1.029 (1.006-1.051) P = 0.012), but not for those treated with artemisinins (OR 0.953 (0.904-1.005) P = 0.078).

Prognostic value of parasite staging
Staging of parasite development on peripheral blood smears provides prognostic information additional to the parasite count itself; the median percentage of ring stages amongst those who survived was 98% and amongst those who died 48% (P = 0.0001). The percentage of the modal   parasite stage was significantly lower amongst those who died than in those who survived (P = 0.0002), suggesting that lower circulating parasite stage synchronicity was associated with death. Using the cut offs of Silamut and White of pigmented stages >10 4 /μL or a parasitaemia of >5 x 10 5 /μL [7], the sensitivity and specificity for predicting death were 84% and 67%, respectively.
died. The post-admission development of complications such as oliguria, seizures or pulmonary oedema and the use of ventilation, lumbar puncture, transfusion and inotropes, were all associated with death (Table 10). Mortality decreased with time, from 9% in 1986-1992 to 7% in 1993-1998 and 6% in 1999-2002 (OR = 0.943 (0.896-0.994), P = 0.030). There was increasing use of artemisinin derivatives over the same periods (0%, 72% and 93% of patients, respectively) and the corresponding percentages of parenteral treatments that were with intravenous/intramuscular artemisinin derivative were 0%, 67% and 93%, respectively. After adjusting for treatment type and study site, there was no trend in mortality over time (OR = 0.961 (0.847-1.091), P = 0.538).

Multiple logistic regression analysis of variables associated with death
All final models were adjusted for treatment with ACT and study site. No interactions between covariates and treatment were significant in any of the models and study site was not significant (P > 0.05). Tables 10 and  11 list variables included in each of the eight final models, all having a maximum of eight variables, consistent with recommendations [31]. Considering the WHO-and AQ-based models, the WHO 1990, WHO 2000, Adapted AQ and Adapted AQ + Pigmented stages, variable sets gave the best predictive power and the areas under the ROC curves (AUROCCs) were not significantly different (Tables 10 and 11, Additional files 2, 3, and 4). Models derived from bedside or bedside + simple laboratory covariates had significantly smaller AUROCCs, when compared to the Adapted AQ model using the common data set (P = 0.008 and 0.016, respectively). Several variables appeared in all four 'best' models and had similar effects across models, such as GCS (OR 0.67-0.73 for increase by one point), serum bicarbonate (OR 0.84-0.86 for increase of 1 mmol/L), plasma lactate (OR 1.20-1.34 for increase of 1 mmol/L), parasitaemia (OR 1.90-4.06 for ten-fold increase) and serum creatinine (OR 1.005-1.006 for increase of 1 μmol/L).
For the MSA score [22] (Additional file 1) using admission variables plus the presence or absence of mechanical ventilation during admission where such ventilation was available (all sites except Sangklaburi) for 635 patients, 516 (81.3%) had an MSA score of 0. Mortality was 2/539 (0.4%) for MSA 0-2, 10/45 (22.2%) for MSA 3-4, 5/12 (41.7%) for MSA 5-6 and 24/39 (61.5%) for MSA ≥7. If ≥5 is taken as the cut off, the PPV for survival was 97.9 (96.4-98.8)%. Among patients who had MSA and BCAM scores calculated (n = 527) the AUROCC for the MSA score in predicting death was 0.97 (0.95-0.98), which was significantly better than that for the BCAM score 0.90 (0.84-0.96) (P = 0.007). The optimal cut off for the data presented here appeared to be <3 rather than <5 as reported [22]; with a cut off of <3 the PPV was 99.6 (98.7-100.0)%, sensitivity 95.1% and specificity 90.4%. Considering all ten models, those based on the MSA score and the Adapted AQ with pigmented parasites had the best predictive power, but AUROCCs for models based on WHO criteria were not significantly lower. Simple rules for classification of severity as a MPI (Table 13, Figure 1), gave sensitivity of 100% and specificity of 82% with co-efficient of 3 rounded to the nearest integer, and sensitivity of 93% and specificity of 92% for a cut off of 4. When compared to the best performing models, based on MSA and Adaptive AQ + Pigmented stages, MPI showed equally good predictive power: In cross-validation, in all runs, the best sensitivity and specificity were obtained for cut offs between 3 and 4 ( Figure 2). For a cut off of 3, the jacknife sensitivity and specificity were 97.1% and 87.1%; for a cut off of 3.5 they were 97.1% and 87.1%; and for a cut off of 4 they were 74% and 94%.

Discussion
The prognosis of severe falciparum malaria has improved markedly since the introduction of parenteral artesunate All estimates are adjusted for treatment with ACT and study site. a the same variables were identified from the WHO (2000) and Adapted AQ models. [21,32]. This large series describing patients with falciparum malaria admitted to hospitals in western Thailand spanned the transition from quinine to artemisinins, and although the data were not from a randomized comparison, mortality was substantially lower in patients who received artemisinin derivatives. Despite the effect of changes in anti-malarial therapy, prognostic indices based on WHO 2000 definitions, and other simpler indices based on fewer variables, provided clinically useful predictions of outcome in Asian adults with severe malaria. Models using variable sets based on WHO 2000 and Adapted AQ with pigmented stages gave the best prediction of mortality, and were comparable to the results based on the MSA score using a smaller sample. Very similar results were also obtained with the MPI based on the most commonly selected variables, GCS, plasma lactate, parasitaemia, serum bilirubin and percentage of pigmented parasites. This will need to be evaluated in independent series of adults with falciparum malaria in similar settings of unstable malaria transmission. This score suffers from the disadvantage, unlike the RCAM score, that determination of 4/5 variables (all except GCS) requires skilled technicians and equipment/consumables and quality assurance that are seldom available where severe malaria is common.
There are at least four limitations of this analysis: specifically, not all patients admitted to the study hospitals were recruited, recruitment criteria varied, there are missing values, and a variety of different doctors reviewed the patients with consequent variability in the nature of both clinical assessment and inpatient management. However, any such differences are likely to result in false negative, rather than false positive associations. The MSA score differs from the others scores discussed, as it is not strictly an admission predictive prognostic score, including mechanical ventilation during hospitalization. Some potential prognostic factors such as haemoglobinuria and abnormal bleeding were too infrequent to allow reliable conclusions to be drawn. Serum bilirubin and plasma lactate were not measured in the South East Asian Quinine Artesunate Malaria Trial (SEAQUAMAT) and therefore the MPI could not be calculated for this dataset [21].  This study differs from that of Hanson [23], which only included patients classified as having severe malaria and was designed to determine which subset of patients could safely be managed without ICU referral. This is reflected in the relatively low mortality in the series described here, as severe malaria was not necessarily a criterion for recruitment. The broader patient population included may well explain the inclusion of parasitaemia as predictive of death in this series and not in that of Hanson [23]. As this series includes many patients without severe disease, the specificity of variables in predicting death may be higher than in series including patients with pre-selected severe disease. However, in busy clinical practice a tool that could define unselected patients admitted to hospital as at risk of death would be valuable.
The wide variation in mortality estimates in this dataset among those with severe malaria (10-18%) and indeed the wide range of estimates of severe disease (6-60%), depending on which definition is used, illustrates the importance of definitions in comparisons between studies, in metaanalyses, and in understanding the host and geographical variability in the presentation and outcome of severe malaria. Various terms, such as 'uncomplicated' and 'mild' , are used to refer to malaria that is not severe. Severe disease is also equated with 'complicated' malaria. To avoid confusion, terminology should be standardized with 'severe' malaria referring to malaria with clinical and/or laboratory features suggesting a clinically significant risk of death (e.g. >5%) despite anti-malarial treatment, and 'uncomplicated' as those without such prognostic features and that the terms 'complicated' and 'mild' are not used.
The present study suggests that, if laboratory tests are available, the history of the illness and the physical examination, apart from GCS and respiratory rate, are relatively unimportant in assessing prognosis in a population of malaria patients. Unlike in other series [13], increased age was not associated with death, except for those treated with quinine. However, as this series did not include children and a smaller percentage (7.7%) were >50 years old, the age range was narrower.
In this series, the mortality of cerebral malaria was increased three-fold if concurrent acidosis and/or renal failure were present and these factors are crucial in predicting death. The ease of identifying patients with MPI A = co-efficients rounded to the nearest integer with cut offs of 3. MPI B = co-efficients rounded to the nearest 0.5 with cut offs of 3. 1 if not available; creatinine >132 μmol/L can be used instead. and cut off of 3 would give sensitivity and specificity of 98%. and 82% for MPI A and 96% and 88% for MPI B. 2 which gives the highest specificity for 100% sensitivity. 3 which gives the "optimal" sensitivity and specificity. Figure 1 Relationship between malaria prognostic index and mortality. Malaria prognostic indices MPI A and MPI B are defined in Table 13.