Syndromic Algorithms for Detection of Gambiense
Human African Trypanosomiasis in South Sudan
Jennifer J. Palmer1*, Elizeous I. Surur2, Garang W. Goch2, Mangar A. Mayen2, Andreas K. Lindner3,
Anne Pittet4, Serena Kasparian5, Francesco Checchi1, Christopher J. M. Whitty1
1Faculty of Infectious and Tropical Diseases, London School of Hygiene and Tropical Medicine, London, United Kingdom, 2Medical Emergency Relief International,
Nimule, South Sudan, 3Medical Mission Institute, Department of Tropical Medicine, Wu ¨rzburg, Germany, 4De ´partement Me ´dico-Chirurgical de Pe ´diatrie, Centre
Hospitalier Universitaire Vaudois, Lausanne, Switzerland, 5Me ´decins sans frontiers, Montre ´al, Canada
Background: Active screening by mobile teams is considered the best method for detecting human African trypanosomiasis
(HAT) caused by Trypanosoma brucei gambiense but the current funding context in many post-conflict countries limits this
approach. As an alternative, non-specialist health care workers (HCWs) in peripheral health facilities could be trained to
identify potential cases who need testing based on their symptoms. We explored the predictive value of syndromic referral
algorithms to identify symptomatic cases of HAT among a treatment-seeking population in Nimule, South Sudan.
Methodology/Principal Findings: Symptom data from 462 patients (27 cases) presenting for a HAT test via passive
screening over a 7 month period were collected to construct and evaluate over 14,000 four item syndromic algorithms
considered simple enough to be used by peripheral HCWs. For comparison, algorithms developed in other settings were
also tested on our data, and a panel of expert HAT clinicians were asked to make referral decisions based on the symptom
dataset. The best performing algorithms consisted of three core symptoms (sleep problems, neurological problems and
weight loss), with or without a history of oedema, cervical adenopathy or proximity to livestock. They had a sensitivity of
88.9–92.6%, a negative predictive value of up to 98.8% and a positive predictive value in this context of 8.4–8.7%. In terms
of sensitivity, these out-performed more complex algorithms identified in other studies, as well as the expert panel. The
best-performing algorithm is predicted to identify about 9/10 treatment-seeking HAT cases, though only 1/10 patients
referred would test positive.
Conclusions/Significance: In the absence of regular active screening, improving referrals of HAT patients through other
means is essential. Systematic use of syndromic algorithms by peripheral HCWs has the potential to increase case detection
and would increase their participation in HAT programmes. The algorithms proposed here, though promising, should be
Citation: Palmer JJ, Surur EI, Goch GW, Mayen MA, Lindner AK, et al. (2013) Syndromic Algorithms for Detection of Gambiense Human African Trypanosomiasis in
South Sudan. PLoS Negl Trop Dis 7(1): e2003. doi:10.1371/journal.pntd.0002003
Editor: John Owusu Gyapong, University of Ghana, Ghana
Received July 13, 2012; Accepted November 28, 2012; Published January 17, 2013
Copyright: ? 2013 Palmer et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits
unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: The research for this study was part-funded by grants from the Sir Halley Stewart Trust (http://www.sirhalleystewart.org.uk/) and the Canadian
Institutes for Health Research, in partnership with the Public Health Agency of Canada (http://www.cihr-irsc.gc.ca/), award no. DPH-88226. The funders had no
role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing Interests: The authors have declared that no competing interests exist.
* E-mail: email@example.com
Found in remote sub-Saharan areas where health systems are
often weak and/or destabilised by armed conflict, human African
trypanosomiasis (HAT, or sleeping sickness) is one of the world’s
most neglected tropical diseases (NTDs). It is caused by
trypanosome parasites that are transmitted primarily through the
bites of infected tsetse flies (Glossina). It is nearly always fatal if
untreated. HAT caused by Trypanosoma brucei gambiense represents
more than 90% of global HAT burden and is endemic in small
geographic foci in 24 countries of west and central Africa .
Humans are assumed to be the main reservoir of infection in these
In the first stage of disease, when parasites are found in the
blood, lymph and organ systems, HAT can be asymptomatic or
involve non-specific symptoms. Patients are considered to be in the
second stage of disease once there is evidence that parasites have
entered the brain and cerebro-spinal fluid (CSF), and it is at this
time that characteristic symptoms are more likely to appear,
involving mental and physical deterioration progressing to death
[2,3]. The natural duration of gambiense HAT is thought to be
almost a year and a half for each stage .
HAT diagnosis is currently not feasible in peripheral primary
health care structures as it requires refrigeration, electricity,
specific equipment and a high level of technical expertise [5,6].
Systematic active screening (AS) of at-risk populations using
laboratory-equipped mobile teams is used to increase access to
treatment and reduce the infectious pool [1,7,8,9], but is
considerably more expensive than passive screening (PS) at static
facilities. Despite calls for an intensification of control activities
coherent with an elimination aim , funding for such activities
on the ground, especially for AS and in the most endemic
PLOS Neglected Tropical Diseases | www.plosntds.org1January 2013 | Volume 7 | Issue 1 | e2003
countries, has recently become worryingly scarce [11,12]. In South
Sudan only 445 people were actively screened for infection in 2010
. In this resource context, it would be useful to develop case-
detection and treatment strategies that can be sustainably
implemented by national and local control programmes. Further-
more, targeting testing to symptomatic patients, who probably
have a higher probability of being HAT cases than the general
population, may be more cost-effective and would reduce the risk
of drug related adverse events among false positives .
One option is to involve non-specialist healthcare workers
(HCWs) in peripheral, first and second-tier primary healthcare
(PHC) facilities in syndromic recognition of suspect cases; these
HCWs are a resource that are often over-looked in vertical HAT
control programmes [15,16]. The WHO cautions against the
exclusive use of signs and symptoms for HAT diagnosis since these
are known to be nonspecific in HAT and their frequency varies
widely between individuals and potentially even geographic
regions . However, a simple syndromic algorithm, if
sufficiently sensitive, could empower HCWs to recognise and
refer suspect HAT patients to a specialised PS facility for
diagnostic HAT testing, and could therefore be a useful tool to
expand case detection. This could also address the problem of
extensive under-diagnosis of HAT at PHC level, which has
commonly been identified as a barrier to passive case-detection
[18,19,20,21]. A predictable drawback of such algorithms,
however, is their low positive predictive value (PPV): HAT
prevalence is typically low (,1–2% in endemic area populations of
the most HAT-affected countries; 2.6%–10.8% in treatment-
seeking populations presenting for PS ), meaning very high
specificity is required to avoid massive over referral. This is
difficult to achieve with high sensitivity.
Two studies [22,23] have investigated the potential for HCWs
to recognize and refer potential cases of HAT based on presenting
symptoms. Boatin et al. (1986) identified a diagnostic scoring
algorithm estimated to be 88% sensitive and 82% specific based on
a comparison of mostly second stage, passively-detected rhodesiense
HAT cases (which feature a different clinical profile and evolution
than in gambiense) and both symptomatic and non-symptomatic
controls in Zambia . Jannin et al. (1993) performed a similar
evaluation for gambiense HAT populations presenting for AS in the
Republic of Congo (RC) and identified an algorithm that was 80%
sensitive with a 20% PPV to detect parasitologically-confirmed
A third study  describes an algorithm instituted in
peripheral facilities in Democratic Republic of Congo (DRC) to
determine the need for laboratory screening, but presents no
information on its accuracy. An additional four studies have
investigated the use of signs and symptoms for classifying disease
stage and/or predicting a fatal prognosis among patients already
confirmed as gambiense HAT cases; none identified a scoring system
sufficiently specific to replace CSF white blood cell counting as a
guide to therapeutic decision-making ([25,26,27] and secondary
analysis of data presented in  by ).
We therefore evaluated the predictive value of these and new
algorithms in a symptomatic, PS service-using population in the
Nimule HAT focus of South Sudan to explore whether any would
be appropriate for application in treatment-seeking populations at
peripheral health facilities where laboratory testing is unfeasible.
The algorithms were designed to be simple to assist HCWs with
limited training. The performance of algorithms was also
compared to referral decisions made by a panel of clinicians with
extensive experience of diagnosing HAT in Africa.
This study was approved by the London School of Hygiene &
Tropical Medicine’s ethical review committee and the Ministry of
Health, Government (now Republic) of South Sudan. All
participants provided verbal informed consent. Verbal consent
was approved for use since most patients presenting to the service
were not literate; receipt of consent was documented by laboratory
staff on data collection forms.
HAT testing and treatment services in the Nimule focus of
Magwi County, Eastern Equatoria State are available at a single
site, Nimule Hospital, supported by the non-governmental
organisation, Merlin (Medical Emergency Relief International)
. Services in this historic focus were re-introduced at the end of
the Sudanese civil war, in 2005. Transmission is thought to have
increased in recent years due to population movements of IDPs
and returnees from neighbouring endemic areas. Small-scale AS
surveys conducted in 2005, 2006 and 2008 revealed an estimated
HAT prevalence of around 1% with the highest prevalence in any
village estimated at 5.8% (Merlin programme data, unpublished).
Community health workers who have received nine months of
formal clinical education are the most common cadre (46%) of
staff involved in patient diagnosis and management in formal
health facilities outside the hospital in the county, followed by staff
with no formal clinical education (39%, mainly nursing assistants
trained on the job), while only 15% are nurses and clinical officers
Collection of patient symptom data
Data on a list of 32 signs, symptoms and epidemiological criteria
(henceforth collectively referred to as ‘symptoms’) were systemat-
ically collected from all patients tested for HAT at Nimule hospital
over a seven month period from October 2009 to April 2010 (see
Table 1). The list of symptoms included those used in other
attempts to create HAT referral algorithms, clinical variables
routinely collected at initiation of HAT treatment in the hospital
Human African trypanosomiasis (HAT or sleeping sickness)
is an almost always fatal disease affecting poor people in
rural, conflict-affected areas of sub-Saharan Africa. It is
difficult to diagnose. Effective treatment exists, but
because diagnostic and treatment services are usually
based only in hospitals, many HAT patients in rural areas
are never detected. Control programmes aim periodically
to extend testing services via mobile teams (active
screening) but their expense and operational issues
severely restrict their use. We explored the predictive
value of different combinations of symptoms that were
present in a treatment-seeking population to identify
people infected with HAT. Through this approach, we
identified a simple four-symptom referral algorithm that, if
replicable, has the potential to identify one HAT patient for
every ten patients referred through subsequent testing. It
would identify most symptomatic HAT patients who seek
treatment, if systematically applied by non-specialist
healthcare workers already working in these areas. As
these types of health workers are rarely included in formal
HAT control efforts, teaching this algorithm also represents
an opportunity to decentralise life-saving knowledge, and
contribute to endemic populations’ long-term empower-
ment and ability to help control this disease.
Symptom Referral in Human African Trypanosomiasis
PLOS Neglected Tropical Diseases | www.plosntds.org2January 2013 | Volume 7 | Issue 1 | e2003
and factors identified as important in local understandings of HAT
from previous qualitative work in the study site . Cases were
defined as (i) positive microscopy on lymph fluid directly or on
blood using capillary tube centrifugation (Woo test) or (ii) positive
CATT on diluted blood serum at dilution 1:16. Patients positive at
dilution 1:8 were considered serological suspects and followed-up
after 3 months. During the study period, first line treatment was
pentamidine for stage 1 and eflornithine for stage 2 cases.
All patient data were collected by laboratory staff (technicians
and assistants without formal clinical training) before HAT test
results were known. Lab staffs were trained to interview patients
and recognise symptoms by the HAT programme manager (ES)
and study coordinator (JP) before data collection. Interviews were
private and patients could be tested without volunteering symptom
information. Symptomatic, HAT-negative patients were advised to
visit the hospital outpatient department for further management.
Patients were only included in the analysis if they presented with at
least one of the symptoms on the list and if complete lab test
outcome data were available. Both naive and previously-treated
patients were included.
Estimation of algorithm performance
The 32 individual candidate symptoms were reduced to a more
computationally and practically manageable set of 13 as follows.
Odds ratios (ORs) were calculated for individual symptoms and
for groups of similar symptoms (e.g., ‘sleep pattern changes’
encompassing ‘daytime sleeping’ and ‘insomnia’). Symptoms that
were significantly associated with being a case or non-case only in
individual analysis were not grouped with others. Symptoms that
were infrequent in the data and/or showed no statistically
significant association with case status were discarded (Table 1
and Table 2). R software, version 2.12 , was used to create all
permutations of 4, 3, 2 and 1-symptom algorithms possible from
the 13 symptoms shortlisted. Pragmatically, we considered that
any algorithm consisting of more than four symptoms would be
inappropriate for peripheral health facilities staffed by HCWs with
Table 1. Presenting symptom data collected and used in algorithm construction.
Individual symptomMethod of ascertainmentRepresentation in algorithms after item reduction
Headache $1 week Patient reportRetained
Back, neck or joint pain Patient report Body pains
Muscle painPatient report Body pains
Fever $1 week Patient reportRetained
Itchy skin Patient reportRetained
Swollen face, legs or arms Patient reportOedema
Weight loss Patient reportRetained
Generally poor state of healthObservationDiscarded
Decrease in appetitePatient report Appetite changes
Increase in appetitePatient reportAppetite changes
Impotence Patient reportDiscarded
No menstruation Patient reportDiscarded
Enlarged lymph nodesExamination Cervical adenopathy
Insomnia Patient report Sleep pattern changes
Daytime sleeping Patient report Sleep pattern changes
Confusion or forgetfulnessPatient report Abnormal behaviour
AggressivenessPatient report Abnormal behaviour
InactivityPatient report Abnormal behaviour
Hallucinations Patient report Abnormal behaviour
Convulsions Patient reportNeurological problems
Difficult speakingObservation Neurological problems
Difficulty walkingObservation Neurological problems
Patient unsteadyObservation Neurological problems
Jerking movementsObservation Neurological problems
Tremor in hands or lipsObservation Neurological problems
Partial paralysisPatient report Neurological problems
Painful tibia/shinExaminationNeurological problems
Treated for malaria/typhoid in last 2 weeksPatient reportRecent malaria and/or typhoid treatment
Works/lives with cows Patient report Patient lives or works with livestock
Works/lives with goats Patient reportPatient lives or works with livestock
Works/lives with sheep Patient report Patient lives or works with livestock
Works/lives with pigsPatient report Patient lives or works with livestock
Symptom Referral in Human African Trypanosomiasis
PLOS Neglected Tropical Diseases | www.plosntds.org3 January 2013 | Volume 7 | Issue 1 | e2003
minimal training. Algorithms were interpreted as indicating
referral for patients presenting with any (as opposed to all) of the
Each algorithm was then tested against the patient data to
compute sensitivity, specificity, PPV and NPV of each, assuming
results of lab testing to be a gold standard diagnosis of HAT. The
Boatin et al. (1986) and Pepin et al. (1989) algorithms were also
tested on the data, as well as three algorithms from the Jannin et
al. (1993) study: the best performing algorithm for detecting all
HAT patients, including those diagnosed on serology grounds
alone (‘Jannin-all’ algorithm) and the two best-performing
algorithms for detecting parasitologically confirmed cases, which
in the original study setting yielded a sensitivity of 84% with a PPV
of 9% (‘Jannin-para2’ algorithm) and 80% sensitivity, 20% PPV
(‘Jannin-para3’ algorithm). Minor deviations from the algorithms
as originally published were made due to the way data were
collected in this study (Table 3).
Comparison with expert clinician performance
Four clinicians with expertise in HAT patient management
were asked to review anonymised patient symptom data and,
blinded to the test outcome, decide whether they would have
Table 2. Crude associations between the 13 symptoms used in algorithm construction and a positive HAT test.
Symptom Cases (%) n=27
Body pains 85.283.5 1.30.4–4.5 0.650
Sleep pattern changes66.7 54.9 1.60.7–3.7 0.238
Headache $1 week55.6 59.3 0.9 0.4–1.90.700
Abnormal behaviour*48.2 26.92.5 1.1–5.60.021
Fever $1 week44.4 32.61.7 0.8–3.6 0.211
Itchy skin*29.69.7 3.91.6–9.7 0.002
Appetite changes22.2 14.7 1.70.6–4.3 0.295
Neurological problems*22.2 4.85.6 2.5–12.4
Oedema14.8 5.33.1 1.0–9.8 0.051
Recent malaria or typhoid treatment14.8 17.50.8 0.3–2.40.724
Weight loss*11.11.8 6.7 1.6–27.20.007
Patient lives/works with livestock 11.110.3 1.10.3–3.70.899
Cervical adenopathy3.7 3.0 1.20.2–9.9 0.834
*Significantly associated with being identified as a case, at p,0.05. Kerendel’s sign (painful tibia) was present in 33.3% of cases and independently significantly
associated with a positive test outcome (individual OR 5.9, p-value ,0.001) but was combined with other more rare symptoms into the larger category ‘neurological
problems’. There were no significant differences in demographic characteristics (age, sex, residency status, location) between cases and non-cases (data not shown). OR:
Odds ratio. CI: Confidence interval.
Table 3. Modified algorithms from other HAT studies tested using Nimule Hospital data.
Symptom Boatin PepinJannin-all Jannin-para2Jannin-para3
Headache $1 week21111
Fever $1 week2-1 1.51.5
Fever unresponsive to anti-malarial-1---
General body pain11---
Neurological problems-- 0.5--
Cervical adenopathy --126.96.36.199
Livestock in compound--111
(Family history of HAT)*--1 1.51.5
Threshold score for referral
Numbers indicate scores attributed to each symptom, if present.
*Data on this symptom were not collected in this study.
Symptom Referral in Human African Trypanosomiasis
PLOS Neglected Tropical Diseases | www.plosntds.org4January 2013 | Volume 7 | Issue 1 | e2003
referred the patient for HAT testing. These experts (ES, AL, AP,
SK) were selected for their substantial experience in direct HAT
patient management in T.b. gambiense-endemic areas (South Sudan,
DRC, RC and Central African Republic). Experts were told only
that patients came from an area with about 1% HAT prevalence
and were asked to make referral decisions (yes or no responses
allowed only) as if they were working in a PHC facility without
HAT testing capacity, one day’s walk from the HAT treatment
centre. The only other information provided on patients was sex,
age (#14 years or .14 years) and whether the patient had ever
been treated for HAT. Accuracy indicators were computed as
above. Qualitative comments from experts about some of the
difficulties they encountered in assigning referral decisions were
taken into account when interpreting subsequent analyses.
The extent of agreement among experts’ referral decisions was
computed so as to establish the level of consensus about what an
appropriate clinical picture for HAT referral might be. Cohen’s
kappa was used to measure inter-rater agreement between pairs of
raters and Fleiss’ kappa between multiple raters . Cases for
which experts unanimously agreed to refer were explored to
identify symptoms that were strongly associated with a unanimous
decision to refer. So as to account for potential confounding in
these associations, a generalised linear model with robust error
variances  was used to estimate these associations by including
all symptoms significant at the 90% confidence level in univariable
analysis into a multivariable model, using a forward stepwise
procedure. The final model contained all symptoms that remained
significantly associated at the 95% level. Age, sex and previous
HAT treatment history were considered as potential confounders
for referral, and were forced into the model to adjust for their
potential effect. All statistical analyses were performed using Stata
software, version 11 (StataCorp, Texas 2009).
Performance of syndromic algorithms
Complete symptom and test outcome data were available for
462/652 (70.9%) patients tested passively for HAT during the 7
month study period. Incomplete symptom data were largely due to
a staff shortage in the hospital laboratory over a 10 week period.
27/462 patients were confirmed as cases, of whom 24 (89%) were
parasitologically confirmed and 24 (89%) were in stage 2, yielding
a prevalence of 5.8% among patients tested.
Out of 14,067 possible candidate algorithms evaluated, more
than half showed a sensitivity greater than 95%, but specificity and
PPV were uniformly low, and the latter remained ,15% even for
algorithms that had a sensitivity as low as 50% (median sensitivity
96.3% (inter-quartile range: 88.9–100.0%), specificity 1.6% (0.2–
6.7%), PPV 5.8% (5.7–5.9%) and NPV 94.9% (87.5–100.0%),
Figure 1). The best-performing algorithms consisted of three core
symptoms (sleep problems, neurological problems and weight loss),
with or without a history of oedema, cervical adenopathy or
proximity to livestock (algorithms 4, 6, 8 and 10 in Table 4,
sensitivity 88.9–92.6%, PPV 8.4–8.7). Algorithm 4 (sleep problems
AND/OR neurological problems AND/OR weight loss AND/
OR oedema) appeared to offer the highest combination of
sensitivity (92.6%) and PPV (8.7%).
The number of presenting HAT symptoms did not appear to
predict infection well: algorithms based on the number of
presenting HAT symptoms (regardless of which) fared worse at
predicting HAT infection than the best-performing algorithm
identified above (e.g. the presence of $4/13 symptoms, regardless
of which, yielded a sensitivity 63% and PPV 8.5; other data not
The Boatin, Pepin and Jannin algorithms also featured low PPV
and, in addition, for the Boatin and Jannin algorithms, low
sensitivity; the Pepin algorithm featured very high sensitivity but
very low specificity (Table 5).
Performance of expert clinicians
When referral decisions were compared to test outcome data,
expert referrers consistently would have referred more cases than
non-cases for screening (Table 6). However, although 3/4 referrers
had high sensitivity, their PPVs were lower than the best-
performing algorithm identified in this study, corresponding to a
higher proportion of patients referred overall ($75% of patients,
as compared to 60%). 62.2% (255/410) of patients were
unanimously referred while only 7.1% of patients in the data
would not have been referred by any expert.
Expert approaches to syndromic HAT referral decisions
Good agreement on expert referral decisions, with an overall
kappa score of 0.56, suggested that experts broadly agreed on what
should constitute an ‘appropriate’ HAT referral, however, agree-
ment on referral of actual cases was in fact quite poor (0.27 for
cases, 0.57 for non-cases) (Table 7) . Multivariable analysis of
symptoms associated with unanimous referral provided insight as
to what experts considered appropriate conditions for referral.
Five symptoms (sleep pattern changes, cervical adenopathy,
neurological problems, recent malaria/typhoid treatment and
abnormal behaviour) retained significance in the final multivar-
iable model at 95% confidence, after adjustment (Table 8). The
only single-symptom algorithm associated with unanimous referral
was ‘sleep pattern changes’. A combination of previous HAT
treatment history and any HAT symptom led to automatic referral
for only one expert referrer. Adults (65.5%) were significantly
more likely to be referred than children under 15 years (43.1%) (p-
value 0.002). There was generally better agreement about what
Figure 1. Receiver operating curve diagram of all candidate
syndromic algorithms evaluated. Each point represents the
sensitivity and 1-specificity of a single algorithm. Ideally, the highest
performing algorithms would be located in the top left corner of the
Symptom Referral in Human African Trypanosomiasis
PLOS Neglected Tropical Diseases | www.plosntds.org5 January 2013 | Volume 7 | Issue 1 | e2003
Table 4. Candidate syndromic algorithms with the 10 highest PPVs and sensitivity $75%.
Sleep pattern change
Headache $ $1 week
Fever $ $1 week
Sx n: The number of symptoms in the algorithm. Y/N: According to the algorithm, the symptom should (Y/yes) or should not (N/no) be present in the patient being referred. When used by HCWs in practice, algorithms containing
more than one symptom should be read as indicating referral if patients present with any (as opposed to all) of the symptoms listed, i.e., symptom A and/or symptom B and/or symptom C and/or symptom D. Patients
referred=the number of patients out of 462 who matched the algorithm being tested.
Symptom Referral in Human African Trypanosomiasis
PLOS Neglected Tropical Diseases | www.plosntds.org6 January 2013 | Volume 7 | Issue 1 | e2003
constituted an appropriate referral between expert referrers and
the Boatin algorithm than with the Pepin or Jannin algorithms.
Using algorithms to identify HAT syndromically
This study suggests that, for users of a PS service in this South
Sudanese focus, algorithms developed for syndromic diagnosis of
HAT in other settings have low sensitivity and poor PPV. Ours is
the first study specifically to examine the potential effectiveness of
syndromic algorithms in a treatment-seeking population and we
present here, a simple four-item syndromic algorithm (sleep
problems AND/OR neurological problems AND/OR weight loss
AND/OR a history of oedema) which had good sensitivity (92.6%)
for detecting HAT in such a context. Under this algorithm, and
given the prevalence observed in Nimule among patients
presenting for testing, about 9 out of 10 referred patients would
be non-cases, corresponding to a relatively low PPV. This PPV
may, however, be considered an acceptable harm for both patients
and health services given the benefit of detecting what is an almost
universally fatal disease. Indeed, such an algorithm may allow
peripheral HCWs to supplement existing passive and (intermittent)
active case finding, and promote integration and strengthening of
HAT services within the overall health system. As with all
algorithms derived from a single patient group it will need to be
tested in independent populations before being recommended for
clinical practice, but we consider it is simple enough to be usable
by HCWs with limited training.
Expert opinion on the potential of syndromic detection of HAT
appears to be unreconciled in the academic literature. Descrip-
tions of how to recognise a case of HAT syndromically abound in
the historic and contemporary HAT literature [3,27,35,36,37,38,
39,40,41,42] and the presence of one particular sign, cervical
adenopathy, was the most important pre-condition for laboratory
testing in AS campaigns over most of the 20th Century. On the
other hand, formal guidance discusses the challenge of operatio-
nalising existing syndromic detection techniques, preferring
instead to advocate widespread use of the more sensitive diagnostic
technologies developed and refined over the last three decades to
avoid the risk of under-detection of a fatal disease [5,6,43]. There
has been much less discussion of the risk of under-detection of
patients who cannot access these technologies directly when,
because of operational realities, use of these technologies is not
widespread, and who might otherwise benefit from a syndromi-
cally-based referral. Perhaps as a result, there are no guidelines
targeted to HCWs working in peripheral facilities to help them
identify suspect patients in these instances.
Other authors have accepted an algorithm with high sensitivity
but low PPV where diagnostic facilities were easily at hand, as in
Pepin et al.’s (1989) study, which equipped all peripheral health
facilities in the district with microscopes; this would however have
greatly increased these HCWs’ workload and in the current South
Sudan context seems unrealistic. Boatin et al. (1986) cautiously
proposed the utility of algorithms with high sensitivity but no PPV
data for application in HCW referrals of rhodesiense HAT, which
occurs sporadically over large areas and thus requires effective
passive case detection. Jannin et al. (1993), however, rejected
algorithms featuring ,10% PPV (albeit with lower sensitivities)
because of the heavy workload implications for HCWs. The
burden placed on patients to travel to a central testing facility,
most of whom would be found negative, was also considered too
great by these authors. However, in our study setting of poorly
accessible PS services and nonexistent AS, an algorithm with 10%
PPV would avert approximately one death for every ten already
treatment-seeking patients who were referred to the testing centre.
Our panel of experts also appeared to implicitly accept this high
rate of false positives, given that they chose to refer a majority of
patients in the case series despite knowing that HAT prevalence
was 1% in the study area.
The reduced accuracy in Nimule of syndromic algorithms
developed in other settings may be due to the lower proportion of
symptomatic patients in those cohorts, as already discussed. It may
also, however, be due to cultural and linguistic differences in
patient perceptions and descriptions of disease. For this reason, it
Table 5. Performance of previously published syndromic algorithms.
Algorithm Patients referred (%)Sensitivity % Specificity %PPV % NPV %
Boatin 244 (52.8)70.448.37.896.3
Pepin 455 (98.5)188.8.131.525.7
Jannin-all362 (78.4)184.108.40.206 96.0
Jannin-para2 163 (35.3)55.666.09.2 96.0
Jannin-para365 (14.1) 29.686.9 12.3 95.2
Table 6. Performance of expert referrers.
referred (%) Chi2p-value
referred (%)Sens % Spec %PPV % NPV %
Referrer 1 20 (74.1)248 (57.0)0.081 268 (58.0) 74.1 43.0 7.596.4
Referrer 226 (96.3) 356 (81.8)0.054382 (82.7)96.318.26.898.8
Referrer 3 26 (96.3)321 (73.8)0.009 347 (75.1)96.326.2 7.599.1
Referrer 4* 26 (100.0)345 (89.8)0.088 371 (80.3)96.389.07.0100.0
*No decision for 52 patients. Pts: patients. Sens: sensitivity. Spec: specificity.
Symptom Referral in Human African Trypanosomiasis
PLOS Neglected Tropical Diseases | www.plosntds.org7January 2013 | Volume 7 | Issue 1 | e2003
would be useful to validate the algorithm more rigorously in an
independent group of patients, first in Nimule and then, if
accuracy is confirmed, elsewhere, to explore its applicability in a
range of health service contexts. In a high HIV burden setting,
three of the four symptoms in this algorithm (neurological
problems, weight loss and oedema) could signal various opportu-
nistic infections associated with AIDS, thus further reducing the
algorithm’s specificity; on the other hand, these patients would also
benefit from hospital referral.
Out of the four symptoms, ‘sleep problems’ was the most
frequently reported in cases and thus made the greatest
contribution to overall sensitivity; however, excessive sleeping
(included in this symptom grouping) was frequently reported by
both cases (63.0%) and non-cases (49.4%), suggesting that HCWs
may face difficulties reliably evaluating its presence in patients. By
contrast, neurological symptoms, weight loss and oedema were
more discriminating for infection, and yet were rarely associated
with HAT by peripheral HCWs interviewed as part of additional
research in Nimule (to be published separately) . Weight loss
and oedema were also not strongly associated with unanimous
expert referral after accounting for confounding, suggesting that to
experts, too, these may be counter-intuitive referral criteria.
Although there appeared to be consensus among them on what
should constitute an appropriate referral, there was poor
agreement on referral decisions for true cases, suggesting that
many of these true cases did not match what could be considered
their ‘consensus case definition’. Alternatively, it may expose
difficulties experts faced in making decisions about these symptoms
with limited information due to the study design.
Five main limitations affect interpretation of this study’s
findings. First, patient symptom reporting may have been subject
to respondent interview bias, including culturally-specific inter-
pretations of some types of symptoms associated with HAT and
other conditions. How patients report HAT symptoms in other
areas may affect this algorithm’s generalisability.
Second, data quality may have been affected by the skill of
laboratory personnel responsible for collecting them. This has
been recognised as a challenge in other studies of HAT
symptomatology  and may be particularly problematic for
recognition of cervical adenopathy [2,22]. It is possible that some
HAT symptoms more prevalent in true cases were under-detected
and therefore had a lower probability of being selected in the final
algorithm. It is debatable whether most referring HCWs in
peripheral facilities would possess a higher level of clinical skill
than the lab attendants trained in this study; if not, the symptoms
in this final algorithm may reflect what is, in fact, most practical, in
this context. Outside of a research setting, hospital lab staff would
not be expected to be involved in syndromic screening since the
more sensitive CATT-WB would be available.
Third, the ‘gold standard’ we used to identify cases in our
analysis was the diagnostic algorithm used in the Merlin HAT
programme, which, as any HAT diagnostic algorithm, is
dependent on the performance of all tests within it and probably
has a true sensitivity of between 85–90% and a PPV of around
90% in PS settings . Little is known about the symptom profile
of false negatives excluded by this diagnostic algorithm so it is
difficult to predict what effect, if any, these exclusions would have
on the sensitivity of the syndromic algorithms we present here.
The Merlin diagnostic algorithm is also probably more sensitive
than gold standard diagnostic algorithms used to evaluate the
syndromic algorithms presented from older studies, making
comparisons less straightforward.
Fourth, our algorithm findings are applicable mainly to
detection of stage 2 patients, since so few patients in stage 1 were
included in the case series. The sensitivity of our algorithms may
be lower in routine peripheral settings if the typical clinical profile
of HAT cases presenting there is less or differently symptomatic,
with proportionately more patients in stage one seeking care.
Finally, our relatively small sample size of 27 cases limits the
precision of our estimates of algorithm sensitivity and PPV.
Consideration of their 95% confidence intervals (e.g., 75.7–99.1%
for sensitivity and 5.7–12.5% for PPV in Algorithm 4) suggests that
they may be considered reasonably accurate but the sample size
Table 7. Kappa scores assessing agreement between all pairs of expert referrers and expert referrers with algorithms from other
Ref 2 Ref 3Ref 4BoatinPepin Jannin-all Jannin-para2Jannin-para3
Ref 10.42 0.53 0.300.330.03 0.09 0.100.10
Ref 2- 0.77 0.850.38 0.070.170.19 0.07
Ref 3-- 0.590.31 0.04 0.130.120.07
Ref 4--- 0.24 0.030.24 0.10 0.03
Ref: Referrer. Kappa scores range from 1 representing complete agreement to 21 representing complete disagreement; a score of 0 represents no more agreement
than would be expected due to chance.
Table 8. Multivariable model of key HAT symptoms
associated with unanimous expert referral, adjusted for age,
sex and previous HAT treatment history (n=407).
Variable RR-adjusted 95% CIp-value
Potential confounding variables
Male sex 1.0 0.9–1.10.703
Patient treated for HAT before 0.80.5–1.2 0.284
Sleep pattern change2.92.3–3.7
Cervical adenopathy1.8 1.2–2.50.003
Neurological problems 1.4 1.3–1.7
Malaria/typhoid treatment 1.31.2–1.5
Abnormal behaviour1.2 1.1–1.40.001
*An additional symptom, body pains, was moderately significant in the final
model (p-value 0.051).
Symptom Referral in Human African Trypanosomiasis
PLOS Neglected Tropical Diseases | www.plosntds.org8January 2013 | Volume 7 | Issue 1 | e2003
may affect the precise ranking of algorithms; practitioners could,
for example, choose to implement or validate any of the top-
performing algorithms identified in this study (algorithms 4, 6, 8
and 10 in Table 4) according to the feasibility of teaching and
Current HAT diagnostic algorithms are too complex for use in
peripheral health structures, and there is very little guidance
available to HCWs working in these areas on when to consider
referring suspected patients to a central testing facility based on
symptoms. This is especially problematic in a context with low AS
coverage like the Nimule focus where most HAT patients in the
periphery are not detected, or in outbreak situations where
funding and capacity for AS is often initially unavailable.
The simple four-symptom referral algorithm identified in this
study has the potential to avert one death through testing and
treatment for every ten patients referred and to identify most
symptomatic HAT patients, if systematically applied. If our
findings can be validated in an independent sample, this algorithm
could represent a useful additional tool for control programmes to
improve case detection in the periphery and promote integration
of HAT services within overall health systems at reasonably low
We would like to thank the study participants, the Nimule hospital
laboratory staff and Merlin project management team, and especially Duku
James Marino for his assistance in data entry and management. We also
thank Prof. Tom Harrison for his comments on an early draft of this paper.
Created template algorithm in R: FC. Wrote first draft: JJP. Wrote redraft:
EIS GWG MAM AKL AP SK FC CJMW. Conceived and designed the
experiments: JJP EIS FC CJMW. Performed the experiments: JJP EIS
GWG MAM AKL AP SK. Analyzed the data: JJP EIS AKL AP SK FC
1. Simarro PP, Jannin J, Cattand P (2008) Eliminating human African
trypanosomiasis: where do we stand and what comes next. PLoS Med 5: e55.
2. Blum J, Schmid C, Burri C (2006) Clinical aspects of 2541 patients with second
stage human African trypanosomiasis. Acta Trop 97: 55–64.
3. Dumas M, Bisser S (1999) Chapter 13: Clinical aspects of human African
trypanosomiasis. In: Dumas M, Bonteille B, Buguet A, editors. Progress in
human African trypanosomiasis, sleeping sickness. Paris: Springer-Verlag
4. Checchi F, Filipe JA, Haydon DT, Chandramohan D, Chappuis F (2008)
Estimates of the duration of the early and late stage of gambiense sleeping
sickness. BMC Infect Dis 8: 16.
5. Chappuis F, Loutan L, Simarro P, Lejon V, Busher P (2005) Options for field
diagnosis of human African trypanosomiasis. Clin Microbiol Rev 18: 133–146.
6. Checchi F, Chappuis F, Karunakara U, Priotto G, Chandramohan D (2011)
Accuracy of five algorithms to diagnose gambiense human African trypanoso-
miasis. PLoS Negl Trop Dis 5: e1233.
7. Cattand P, Jannin J, Lucas P (2001) Sleeping sickness surveillance: an essential
step towards elimination. Trop Med Int Health 6: 348–361.
8. Meda H, Pepin J (2001) The epidemiology and control of human African
trypanosomiasis. Adv Parasitol 49: 71–132.
9. Simarro PP, Diarra A, Postigo J, Franco JR, Jannin JG (2011) The human
African trypanosomiasis control and surveillance programme of the World
Health Organisation 2000–2009: The way forward. PLoS Negl Trop Dis 5:
10. WHO (2007) Report of a WHO informal consultation on sustainable control of
human African trypanosomiasis. Geneva: World Health Organisation. WHO/
11. Jannin JG, Simarro PP, Franco JR (2011) A11 Progress in control and
elimination of human African trypanosomiasis, 2010. In: Choffnes E, Relman D,
editors. National Academies Press.
12. Yun O, Priotto G, Tong J, Flevaud L, Chappuis F (2010) NECT is next:
implementing the new drug combination therapy for Trypanosoma brucei gambiense
sleeping sickness. PLoS Negl Trop Dis 4: e720.
13. Ruiz-Postigo JA, Franco JR, Lado M, Simarro PP (2012) Human African
trypanosomiasis in South Sudan: how can we prevent a new epidemic? PLoS
Negl Trop Dis 6: e1541.
14. Inojosa WO, Augusto I, Bisoffi Z, Josenado T, Abel PM, et al. (2006) Diagnosing
human African trypanosomiasis in Angola using a card agglutination test:
observational study of active and passive case finding strategies. BMJ 332: 1479.
15. Mwanakasale V, Songolo P (2011) Disappearance of some human African
trypanosomiasis transmission foci in Zambia in the absence of a tsetse fly and
trypanosomiasis control program over a period of forty years. Trans R Soc Trop
Med Hyg 105: 167–172.
16. Bilengue CM, Meso VK, Louis FJ, Lucas P (2001) [Human African
trypanosomiasis in the urban milieu: the example of Kinshasa, Democratic
Republic if the Congo, in 1998 and 1999]. Med Trop (Mars) 61: 445–448.
17. WHO (1998) Control and surveillance of African trypanosomiasis: report of a
WHO expert committee. Geneva: WHO.
18. Hasker E, Lumbala C, Mbo F, Mpanya A, Kande V, et al. (2011) Health
care-seeking behaviour and diagnostic delays for Human African Trypano-
somiasis in the Democratic Republic of the Congo. Trop Med Int Health 16:
19. Bukachi SA, Wandibba S, Nyamongo I (2009) The treatment pathways followed
by cases of human African trypanosomiasis in western Kenya and eastern
Uganda. Ann Trop Med Parasitol 103: 211–220.
20. Odiit M, Shaw A, Welburn SC, Fevre EM, Coleman PG, et al. (2004) Assessing
the patterns of health-seeking behaviour and awareness among sleeping-sickness
patients in eastern Uganda. Ann Trop Med Parasitol 98: 339–348.
21. Kovacic V (2009) Health seeking behaviour in relation to sleeping sickness
(Human African trypanosomiasis) in West Nile, Uganda [MPhil]. Oxford:
University of Oxford.
22. Jannin J, Moulia-Pelat J, Chanfreau B, Penchenier L, Louis J, et al. (1993)
Trypanosomiase humaine africaine: etude d’un score de presomption de
diagnostique au Congo. Bull World Health Organ 71: 215–222.
23. Boatin B, Wyatt G, Wurapa F, Bulsara M (1986) Use of symptoms and signs for
diagnosis of Trypanosoma brucei rhodesiense trypanosomiasis by rural health
personnel. Bull World Health Organ 64: 389–395.
24. Pepin J, Guern C, Milord F, Bokelo M (1989) [Integration of African human
trypanosomiasis control in a network of multipurpose health centers]. Bull
World Health Organ 67: 301–308.
25. Bertrand E, Serie F, Kone I, Rive J, Compaore P, et al. (1973) Symptomatologie
ge ´ne ´rale de la trypanosomiase humaine africaine au moment du d’epistage. Med
Afr Noire 20: 303–314.
26. Edan G (1979) Signes cliniques et biologiques des trypanosomiases a T. gambiense
vues au stade d’atteinte meningo-encephalitique. Med Trop (Mars) 39: 499–507.
27. Antoine P (1977) Etude neurologique et psychologique de malades trypanosomes
et leur evolution. Ann Soc Belge Med Trop 57: 227–247.
28. Boa Y, Traore M, Doua F, Kouassi-Traore M, Kouassi B, et al. (1988) Les
differents tableaux cliniques actuels de la trypanosomiase humaine africaine a Tb
gambiense: analyse de 300 dossiers du foyer de Daloa, Cote-D’Ivoire. Bull Soc
Pathol Exot Filiales 81: 427–444.
29. Kegels G (1997) ‘‘Vertical analysis’’ of human African trypanosomiasis. Studies
in health services organisation and policy. Antwerp: ITG Press.
30. Palmer J (2012) Utilisation of human African trypanosomiasis passive screening
services in post-conflict South Sudan [PhD]. London: London School of
Hygiene & Tropical Medicine.
31. R Development Core Team (2011) R: A Language and Environment for
Statistical Computing. 2.12 ed. Vienna, Austria: R Foundation for Statistical
32. Fleiss J (1981) Statistical methods for rates and proportions. New York: Wiley.
33. Zou G (2004) A modified poisson regression approach to prospective studies with
binary data. Am J Epidemiol 159: 702–706.
34. Kirkwood B, Sterne J (2003) Essential Medical Statistics. Oxford: Blackwell
35. Apted F (1970) ‘‘Chapter 35: Clinical manifestations and diagnosis of sleeping
sickness’’. In: Mulligan H, editor. The African trypanosomiases. London:
George Allen and Unwin Ltd.
36. Giordano C, Clerc M, Doutriaux C, Doucet J, Nozais J, et al. (1977) Le
diagnostique neurologique au cours des differentes phases de la trypanosomiase
humaine africaine. Ann Soc Belge Med Trop 57: 213–225.
37. Burri C, Brun R (2003) Chap 73: Human African trypanosomiasis. In: Cook G,
Zumla A, editors. Manson’s Tropical Diseases. 21st ed. London: Elsevier
Sciences. pp. 1303–1323.
38. Tooth G (1950) Studies in mental illness on the Gold Coast. Colonial Research
Publication No. 6. London: His Majesty’s Stationery Office.
39. Gelfand M (1947) Transitory neurological signs in sleeping sickness. Trans R Soc
Trop Med Hyg 41: 225–258.
40. Stich AH (2012) African trypanosomiasis. In: Mabey D, Parry E, Gill G, et al,
editors. Principles of medicine in Africa. 4th ed. Cambridge: Cambridge
Symptom Referral in Human African Trypanosomiasis
PLOS Neglected Tropical Diseases | www.plosntds.org9January 2013 | Volume 7 | Issue 1 | e2003
41. Low G, Castellani A (1903) Report on sleeping sickness from its clinical aspects. Download full-text
Reports of the Sleeping Sickness Commission of the Royal Society 2: 14–63.
42. Lambo TA (1966) Neuro-psychiatric syndromes associated with human
trypanosomiasis in tropical Africa. Acta Psychiatry Scandinavia 42: 474–484.
43. Corty J (2011) Chapter 7: Human African trypanosomiasis. In: Bradol J, Vidal
C, editors. Medical interventions in humanitarian situations: The work of
Medecins Sans Frontieres: MSF-USA.
44. Tshimungu K, Okenge L, Mukeba J, Kande V, De Mol P (2009) Caractristiques
epidemiologiques, cliniques et sociodemographiques de la trypanosomiase
humaine africaine (THA) dans la region de Kinshasa, Republique democratique
du Congo. Sante 19: 73–80.
Symptom Referral in Human African Trypanosomiasis
PLOS Neglected Tropical Diseases | www.plosntds.org10 January 2013 | Volume 7 | Issue 1 | e2003