Optimising mHealth helpdesk responsiveness in South Africa: towards automated message triage

Acquisition and demographics of incoming messages

Individuals attending public health facilities in South Africa can send a text message via SMS to the helpdesk at any time to communicate directly with NDoH staff. Messages received between 27 October 2016 and 17 July 2017 were downloaded from the District Health Information System (DHIS2) for this analysis. Responses to these messages by helpdesk staff were also downloaded, and incoming messages were paired to the corresponding responses for subsequent analysis.

A total of 65 768 messages acquired during the aforementioned window were available for analysis. In total, 49 300 (75.1%) of these were questions, 9189 (14.0%) were message switch requests, 3121 (4.8%) were compliments, 2561 (3.9%) pertained to prevention of mother-to-child transmission of HIV (‘PMTCT’), 560 (0.9%) were opt-out requests, 361 (0.5%) were complaints, 351 (0.5%) were language switch requests, 109 (0.2%) were spam and 126 (0.2%) could not be classified.

The helpdesk users sending these messages ranged from 13 to 52 years old, with an average age of 27.8±6.15 years and an IQR of 9.2 years. The largest number of messages were received from KwaZulu-Natal (28.1%) followed by Gauteng (22.1%), Limpopo (12.9%), Mpumalanga (9.4%), Eastern Cape (8.8%), North West (7.5%), Western Cape (5.7%), Free State (4.5%) and lastly Northern Cape (1.0%). The majority of registered users (50.2%) were recorded as possessing a South African National ID card, and approximately a third (33.0%) opted to receive messages in English. Gestational age averaged 19.5±8.4 weeks with an IQR of 12.1 weeks.

Message handling: random sample

To characterise the response to these messages, a random sample of 481 English-language messages was drawn from the complete message set. This sample size was chosen to obtain desirable (<4%) margins of error for the scoring categories described in the next section.

Helpdesk responses were characterised by two measures: the appropriateness of the response and the time delay between the incoming message and its associate response. Response appropriateness was manually scored as follows: a score of 0 (none) indicates no reply was given, 1 (incorrect) indicates the reply did not make sense or otherwise seemed to be in error, 2 (suboptimal) indicates an inappropriate or incomplete response, and all other responses were scored as 3 (satisfactory). Time delay (or response time) was measured as the difference between timestamps associated with a query and its corresponding reply. If more than one reply was sent, the earliest timestamp was used in the calculation.

Among the randomly sampled messages (n=481), the age of users ranged from 15 to 44 years, with an average age of 26.6. While all provinces were represented, the majority of messages were sent from users in Gauteng (116 messages, 24.1%) and the least from Northern Cape (4 messages, 0.8%).

A total of 391 messages (81.3%±3.5%) were scored as optimal, 80 (16.6%±3.3%) as suboptimal and 10 (2.1%±1.3%) as incorrect. Median response time was 4.0 hours among all sampled responses, 4.5 hours among those graded as optimal, 2.1 hours among those graded as suboptimal and 17.2 hours among those graded as incorrect. Figure 1 shows response times for these categories; a logarithmic scale has been used due to the presence of extreme outliers. The difference in response time between optimal and suboptimal responses was not statistically significant (P=0.07). With only 10 messages graded as incorrect, differences in response time between this group and the others were not statistically significant (P>0.3).

Request permission

Identifying mistreatment of women

Messages were screened for the presence of keywords pertaining to mistreatment (table 1). Keywords were identified based on frequent typologies of abuse reported in the literature.3 5 6 8–10 Those matching any of the keywords were flagged for examination, and the presence of one of the typologies listed in the table was manually confirmed. Messages were further divided according to whether the mistreatment occurred at a health facility or at home; the former were categorised as facility-based mistreatment and the latter as partner and family violence. All messages identified could be reliably assigned to one of these categories, so no other categories were needed. Additionally, all messages labelled as a complaint by helpdesk staff were manually examined for the presence of facility-based mistreatment or partner and family violence. Those categorised as facility-based mistreatment were assigned to one of the typologies in table 1.

Out of the full message set (n=65 768), 32 reports of facility-based mistreatment and 39 reports of partner or family violence were identified. A further breakdown of these counts along with illustrative messages and responses is given in table 2. Additionally, 19 messages reporting poor service at facilities were found.

Our ability to detect these reports of mistreatment and partner and family violence via keyword supports the feasibility of flagging high-priority messages. When the keywords in table 1 are present, it is likely that the message pertains to mistreatment. This finding could not be more rigorously evaluated using a classifier in the current work, partly because rates of mistreatment reporting were low, but more importantly because there was no ‘ground truth’ regarding mistreatment. In other words, we had no independent method of verifying whether mistreatment was reported in the 65 768 messages available. Nevertheless, this result provides preliminary evidence that keyword-matching can form the basis of a message triage system.

Message handling: mistreatment of women

In the majority of the facility-based mistreatment cases (72%; 23 of 32), the message was acknowledged as a complaint using the following reply: “Thank you for sending in your complaint. We have taken note of it and will log the complaint with the Department of Health and your facility.” In three cases, the helpdesk directed the woman back to the clinic where she had been mistreated.

Response appropriateness was compared between both facility-based mistreatment and partner and family violence and the random sample (n=481) via χ² test; both message sets were found to be similar to the random sample. There appears to be a slight but not significant trend towards better handling of partner and family violence, with 34 of 39 scored as optimal (87.2%) compared with 391 of 481 (81.3%) in the random sample (P=0.182); and poorer handling of complaints about facilities (poor service or facility-based mistreatment), with 38 of 51 (74.5%) scored as optimal (P=0.183). Table 3 provides counts of the scores in each message set, along with their expected values calculated during χ² testing. The latter shows counts that would be expected if response handling were the same between sets.

Response time was also compared between both facility-based mistreatment and partner and family violence and the random sample via Kolmogorov-Smirnov test. Figure 2 shows the distribution of response time between these three groups. Median response times were 13.1 hours for facility-based mistreatment, 2.0 hours for partner and family violence, and 3.7 hours in the random sample. The differences in response time between facility-based complaints and the random sample did not reach statistical significance (P=0.074). Response time was similar between partner and family violence and the random sample (P=0.566).

Request permission

Age was tested for normality by Shapiro-Wilk test, then compared pairwise between groups by Mann-Whitney U test. This analysis showed that age was not normally distributed in our sample (P<0.001) and was similar between mothers from the random sample and those with facility-based complaints (P=0.267) and experiencing partner and family violence (P=0.516). Differences among provinces were compared by χ² test, but no statistically significant association between mistreatment reporting and province was detected (P=0.169).

The need for triage is supported by the presence of these reports of mistreatment as well as their handling by helpdesk staff. While these messages are uncommon, proper identification and handling is critical to provide the best possible support for helpdesk users and also for the purposes of health sector oversight and accountability. Results did not conclusively demonstrate that handling of these messages is inferior to the handling of others, but neither is it superior. With a median response time of 8.6 hours and 18 non-optimal responses among the 90 total reports of mistreatment, there is certainly room to improve. Given the large volume of messages received by the limited number of helpdesk staff, triage is needed to improve this performance.

Limitations

Our keyword-based approach, while appropriate under the circumstances, is not guaranteed to identify all messages reporting mistreatment. The list of keywords is not exhaustive and does not account for misspellings and other typos or character errors. Manual inspection could have been used to more reliably identify mistreatment, but was prohibitively labour intensive on a message set of this size. Further, the keywords are all in English; a major limitation was our inability to identify mistreatment not reported in English in this analysis.

Messages reporting mistreatment were limited in number, with 32 facility-based mistreatment and 39 partner and family violence messages identified. This limited sample size prevented us from drawing definitive conclusions about the handling of these messages (response time and appropriateness) as well as the demographic determinants of mistreatment (age and province). While several trends have been identified, they did not reach statistical significance.

Classifying text via naïve Bayes

To explore the feasibility of automated triage, a naïve Bayes classifier was trained to label incoming messages. Naïve Bayes is a simple, scalable and robust classifier with a long history of use in text classification.11 In brief, naïve Bayes is a probabilistic classifier that simplifies the likelihood of a given class using the naïve Bayes assumption, which stipulates that features—in this case, individual word occurrences—are conditionally independent given the class label. This assumption may be written as follows:

In this equation, the W_i represent counts of the n possible words of interest, and C is the class label—in our case, the message label. Having made this assumption, the terms P(W_i|C) may be calculated simply as the frequencies of word occurrences for each of the labels. New messages may then be classified based on Bayes’ rule as:

For this application, the message labels C belong to one of the following 10 categories: ‘Question’, ‘Message Switch’, ‘Compliment’, ‘PMTCT’, ‘Opt Out’, ‘Complaint’, ‘Language Switch’, ‘Spam’ and ‘Unable to Assist’. Words are counted in each message, making ours a multinomial model.11 In addition to its simplicity and history of success in similar applications, the naïve Bayes approach is equally simple to implement in a multiple-language setting, giving it an edge over newer, more sophisticated classifiers for the current application. Indeed, naïve Bayes should perform equally well even when multiple languages are present in a single message, which is common in this dataset.

The 65 768 available, labelled helpdesk messages were divided into a training set (n=53 232) used to train the classifier and a test set (n=12 536) used to evaluate its performance. Individual words were identified by their word stems; for example, ‘training’ and ‘train’ count as the same word. The classifier was trained to replicate the labels used by NDoH staff based on the word stems present in a message. Following training, performance was evaluated on the test set using a confusion matrix of message labels correctly and incorrectly predicted by the classifier. Sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) were then calculated for each label.

Classifier performance

Overall accuracy for the naïve Bayes classifier was 85.4%, with 95% CI ranging from 84.8% to 86.0%, and Cohen’s kappa coefficient equal to 0.556. Classification accuracy was conclusively superior to the no-information rate of 82.4% (P<10⁻¹⁰). The confusion matrix for the naïve Bayes classifier is presented in table 4; this matrix summarises performance by cross-tabulating NDoH-assigned labels with our classifier’s predictions. Sensitivity, specificity, PPV and NPV for each label are presented in table 5.

Specificity was lowest for ‘Question’, the most common label, but ≥98% for all other labels. Similarly, NPV was ≥97% for all labels other than ‘Question’. Sensitivity and PPV were less consistent. Questions, message switch requests, compliments and PMTCT had sensitivity and PPV both >50% and as high as 93%. On the other hand, none of the small number of spam and unable-to-assist messages were correctly identified by the classifier—most were labelled as ‘Question’, as shown in table 4—resulting in sensitivity and PPV of 0 for those labels.

These results—along with successful identification of mistreatment—strongly suggest that helpdesk triage is feasible. As hypothesised, the naïve Bayes classifier had high sensitivity for all labels except the most common one, ‘Question’. While higher sensitivity is desirable, high specificity is most essential for the purposes of triage: it ensures that staff would not be inundated with messages that have been incorrectly labelled as high priority. When specificity is low, high-priority labels would become less useful to helpdesk staff due to comparatively larger number of false positives. Varying sensitivity between labels implies that in some cases, messages deserving high-priority status will not be successfully identified; however, this is no worse than the current system, under which no messages are flagged. PPV is fairly high for several labels other than ‘Question’, showing that when these labels are assigned, they are meaningful.

More generally, these results add to a large body of evidence supporting use of the naïve Bayes classifier in text classification11 and demonstrate that a ‘bag of words’ approach is effective for SMS labelling and triage. As the phrase ‘bag of words’ implies, this approach considers messages only as a set of their constituent words—or in this case, word stems—without considering syntax or word order. Similar to the keyword-matching used to identify mistreatment, naïve Bayes relies on the presence or absence of specific indicator words to guide classification. If a message contains the word ‘beat’ or ‘thank’, for instance, it is likely to be a report of violence or a compliment, respectively.

Because it does not rely on word order or syntax, this method of classification can perform well in any language or multiple languages. This is an important advantage in the current setting. However, good performance across the 11 languages supported by the helpdesk would require adequate training data in each of them, which could prove difficult to obtain. More generally, this method can be used in any two-way communication platform able to consistently identify individual words.

Taken together, these analyses show that the handling of high-priority messages could be improved, and that a simple classifier is capable of automatically flagging important but infrequently occurring messages. Thus, automated triage could feasibly be used to identify such messages for priority handling by NDoH staff, likely leading to meaningful helpdesk improvement.

Limitations

Our method does not take advantage of word order or syntax, but instead understands each message as a ‘bag of words’. While we have cited this as an advantage in the current application, particularly given the multiple languages present in our message set, a syntax-aware approach might yield better results. The conditional independence assumption of naïve Bayes is another well-known limitation, but one that has not hindered its effective use in many text classification applications.

Importantly, while our results support naïve Bayes as the basis for automated triage, training such a system would first require manual identification of high-priority topics among a large set of messages. This burden could be reduced to some degree using expert knowledge and/or active learning.

As an exploratory analysis, this work is only a first step in the development of automated triage, which would require continued development of a triage system; its integration into the helpdesk platform; usability testing to integrate the triage process into existing helpdesk workflows; and ultimately an evaluation of its effectiveness via prospective trial. Nevertheless, the current analysis reveals that automated triage is a promising means to streamline and improve the helpdesk even as the user-base continues to grow. This in turn helps to ensure the helpdesk remains an effective mechanism for empowering and mobilising women and for improving the quality of care provided and promoting health systems’ accountability.