Dengue fever (DF) is one of the most common arthropod-borne viral diseases worldwide [1], especially in South East Asia, Africa, the Western Pacific, and the Americas [2,3].

However, there is no accurate and speedy diagnostic screening test for DF at an early stage, as its signs and symptoms—for example, fever, headache, and myalgia—are similar to those of other illnesses [4-6]. Some studies [4,5] that used a univariate analysis report that the presumptive diagnosis of DF is imprecise. Multivariate logistic regressions also do not significantly distinguish patients with dengue from those with other febrile illnesses [7]. The multivariate discrimination analyses reported sensitivity and a specificity 0.76 and an area under the receiver operating characteristic (ROC) curve (AUC) of 0.93, but costly laboratory tests (Dengue Duo Immunoglobulin M and Rapid Strips, Panbio, Queensland, Australia) [8-11] were needed before DF was serologically confirmed.

DF symptoms are usually assessed using a dichotomous (ie, absent vs present) evaluation. The dependent variable (DF⁺ vs DF⁻) predicted using independent evaluations with a weighted summation score is more accurate than that predicted using simple evaluations with an unweighted summation score. So far, there has been no published study that has reported using the specific sequence of symptoms reported or observed in specific patients suspected of having DF. All published studies to date still report results using only a standard group of symptoms with an unweighted summation score, and they merely apply their results to a general group of patients who might have DF.

The nonparametric HT fit statistic has been used in education and psychometrics to identify aberrant test respondents [12,13]. It is a transposed formulation of a scalability coefficient for items (eg, symptoms in this study), and it is the best among 36-person fit statistics for detecting abnormal behaviors [14].

In this study, we used the HT coefficient combined with weighted and unweighted variables to examine whether these combinations provide a valid and reliable approach for the early detection of DF in children.

The sample of 177 pediatric patients (≤16 years old; DF⁺: 69; DF⁻: 108) was the same as in our previous paper [8] (see data in Multimedia Appendix 1). Guided by the literature [5-7], we collected 19 DF-related clinical symptoms from the patients’ medical records to develop the initial set of items—designated as 0=“absent” or 1=“present”—to screen for DF infection: (1) personal history of DF, (2) family history of DF, (3) mosquito bites within the previous 2 weeks, (4) fever ≥39°C, (5) biphasic fever, (6) rash, (7) petechiae, (8) retroorbital pain, (9) bone pain (arthralgia), (10) headache, (11) myalgia, (12) abdominal pain, (13) anorexia, (14) occult hematuria, (15) stool occult blood, (16) cough, (17) sore throat, (18) soft (watery) stool, and (19) flushed skin. Data from these patients’ charts were obtained and approved by the Research Ethics Review Board of the Chi-Mei Medical Center.

HT is defined for the persons of a dichotomous dataset with L items (in columns) and N persons (in rows) [12-14], where X_ni is the scored (0,1) response of person n to item i, and P_n=S_n/L. Here, S_m is the raw score for person m, and S_n is the raw score for person n.

HT is the sum of the covariances between person n and the other persons divided by the maximum possible sum of those covariances so that the range of HT is from −1 to +1, see formula (1) in Figure 1. When the responses by person n are positively correlated with those of all the other persons, then HT (n) will be positive. In contrast, when the responses by person n are negatively correlated with those of all the other persons, then HT (n) will be negative. When person n ’s responses are random, HT (n) will be close to zero [11]. We hypothesized that DF⁺ patients have different HT coefficients than DF⁻ patients. All DF⁺ group members were sequenced to the DF⁻ group members to obtain an HT coefficient using formula (1) in Figure 1.

All symptoms were examined by the probability of Type 1 error using the following 3 steps in Figure 2 to determine predictor variables. First, each symptom was separately examined by the univariate approach using a Chi-square test and logistic regression, respectively, for identifying a significant association with DF. Second, 2 models (ie, the univariate and the multivariate approaches) were investigated for determining valid predictor variables associated with DF when the probability of Type 1 error was less than .05. Third, the predictor variables were used in a weighted combination for discriminating patients suspected with dengue virus infection.

The efficacy of 3 models (A, B, and C) for detecting dengue fever was examined: (1) A comparison was made using univariate logistic regression in Model A to examine effects through the AUC, yielded by unweighted (ie, summed item) scores, weighted (ie, logistic regression) scores, and HT coefficients, respectively. (2) Multivariate logistic regression with the 3 aforementioned factors combined was used in Model B. (3) After selecting the significant variables in Model B, the combined predictive variables were analyzed using multivariate logistic regression in Model C to obtain effective weighted coefficients. (4) Finally, we wanted to use a single continuous variable yielded by the combined predictive variables in Model C to compare the AUC with the counterparts in Model A and C.

Moreover, we provide the F-measure for evaluating the predictive effect [15], which is calculated by following equations: precision=True Positives/(True Positives+False Positives); recall=True Positives/(True Positives+False Negatives); F-measure=(2×precision×recall)/(precision+recall).

SPSS 15.0 for Windows (SPSS Inc) and MedCalc 9.5.0.0 for Windows (MedCalc Software) were used to calculate (1) the probability of false positives (Type 1 error) using a Chi-square test and logistic regression, (2) Youden J index (the higher, the better), AUC, sensitivity, specificity, and the cutoff point at maximal summations of specificity and sensitivity, (3) correlation coefficients among variables of unweighted, weighted, and HT scores.

A total of 69 pediatric patients clinically diagnosed with DF and 108 pediatric patients with no evidence of DF infection were included in this study (Table 1). A Chi-square test and logistic regression analyses showed that only 6 symptoms (family history, fever ≥39°C, skin rash, petechiae, abdominal pain, and weakness) were significant for assessing the likelihood of DF (Table 2).

Comparisons of the AUCs for the 3 study models (A, B, and C) showed that the weighted variable (derived by the Logistic regression) and the HT coefficient could be jointly used for predicting DF risk using equation (2):

The risk probability can be computed using the transformed formula 3:

where logit denotes a unit of log odds.

A cutoff point of >–0.68 (P=.34) was determined using the combined predictive variables in Model C: sensitivity=0.91, specificity=0.76, AUC=0.88, and the highest F-measure=0.82 (see Figure 3 and Table 3). Predictive power was better: specificity was 10.2% (ie, 84.30–74.10, shown in Table 3) higher than when using traditional logistic regression, that is, the independence variable=sum (weighted score for each symptom x the respective symptom response, 1 or 0, predicting the dependence variable, 1 or 0 for DF). Even if AUC using the HT coefficient was slightly lower (0.72) than when using the unweighted (0.84) and the weighted (0.87) variables (Table 3), and the HT coefficients related to the weighted and unweighted scores were 0.26 and 0.22, respectively, the weighted score had a higher correlation coefficient than the unweighted score to the HT coefficients, and the combined strategy of Model C or the single continuous variable yielded by the combined predictor variables (Table 3) are verified and available for use in practice. More importantly, the sensitivity is more critical than the specificity in clinical settings, as we would not miss any 1 case with fatal diseases.

A snapshot on a mobile phone responding to questions (Figure 4, top) was generated, and the results for assessing whether the patient has DF (Figure 4, bottom) were determined, which indicated that patients suspected of having DF could directly scan the Quick Response Code to obtain their DF logit scores (or the risk probability) and examine whether these 6 symptoms are useful for predicting a high DF risk (>−1.03 logits or P ≥.26=exp(−1.03 logits)/(1+exp(-1.03 logits)). Interested readers are recommended to see the demonstration in Multimedia Appendix 2 using a MP4 video to display.

We found that using the HT coefficient yielded predictions that were 10.2% more specific (ie, 84.30–74.10, shown in Table 3) than those of traditional logistic regression. The HT index is promising when the patient sequence symptom pattern is compared with the DF⁺ group to detect dengue fever in children. It can be combined with the weighted summation score to jointly predict the DF risk and then report that risk on mobile phones.

The HT coefficient has been used in education and psychometrics to identify aberrant test respondents [12-14]. Although some have used item response theory fit statistics (eg, outfit mean square error >2.0) to select abnormal responses that indicate cheating, careless responding, lucky guessing, creative responding, or random responding [16], our literature review revealed no published papers that reported using the HT coefficient in medical settings, especially for detecting individual aberrant response patterns different from the study reference sample, or, like this study, identifying the DF risk by comparing their sequence symptom pattern with that of the DF⁺ group.

A diagnosis of DF is usually confirmed by 3 steps: (1) observing DF-related symptoms, (2) testing laboratory data, such as white blood cells and platelets, and (3) serologically verifying DF using dengue Immunoglobulin M and Immunoglobulin G antibodies, polymerase chain reaction analysis, and virus isolation tests [8]. The latter 2 are relatively expensive. It is needed to develop a self-assessment approach (eg, scanning Quick Response Code, responding questions, and obtaining the DF risk on his/her smartphone), (1) helping patients for consultation at an earlier stage and (2) prompting doctors for sampling patient laboratory data when his/her DF risk reaches a cut point of P=.26 (=exp(−1.03 logits)/(1+exp(−1.03 logits)).

We found that the weighted score was a better predictor than the unweighted score (see Model A and Model B in Table 3). However, we still see so many scales in a medical setting using unweighted summation scores to determine the presence or absence of disease. Along with the mobile phones popularly used in the technical age, the way of obtaining the DF risk on mobile phones using the combined HT coefficient and weighted scores is available and worth recommending to health care providers to use for detecting the risk for DF.

This study has some limitations. First, the DF cut point based on the symptoms of this study sample might be biased toward that population. Moreover, we did not remove abnormal data when the HT coefficient was less than the critical value of 0.22, which best identifies aberrantly responding examinees [14]. Second, although the sample size was small, using the HT coefficient combined with the AUC yielded highly accurate discriminatory screening. However, this finding requires confirmation in prospective studies of other regions with a substantial incidence of DF. Third, the study sample size (=177) is too small to make the inference reliable and supportable. More DF patients collected in a study are required to be considered in the discernable future. Particularly, artificial intelligence (AI) has become increasingly prevalent in recent years.

Analyzing 6 simple symptoms using logistic regression is useful and valid for the early detection of DF risk in children. Combining the HT coefficient with the weighted score yields a prediction that is 10.2% more specific than that yielded by traditional logistic regression. A self-assessment app using patient mobile phones is available to help people suspected of having DF, and it might eliminate the need for costly and time-consuming laboratory tests.