A Clinical Prediction Rule for the Severe Acute Respiratory Syndrome

IN RESPONSE:

Ma and colleagues' results after applying our clinical prediction rule confirm the generalizability of the algorithm beyond Hong Kong to another urban population affected by SARS. This satisfies the call by the U.S. Centers for Disease Control and Prevention to externally validate our decision rule to assess its clinical effectiveness for potential future outbreaks (1).

For cross-validation purposes, we applied the Taiwanese rule (2), previously published by Ma's group, to our sample: 2649 emergency department attendees at the Prince of Wales and United Christian Hospitals, which provided care for the 2 largest outbreak clusters in Hong Kong. The Table shows the performance indices associated with both the Hong Kong and Taiwanese decision rules. In the base-case analysis, Ma and colleagues' symptom scoring system achieved a sensitivity of 0.513 and a specificity of 0.583, with an area under the receiver-operating characteristic curve of 0.57; for the clinical scoring system, these values were 0.481, 0.764, and 0.65, respectively. If this rule were applied during a large-scale outbreak, such a low sensitivity and similarly suboptimal performance indices would be unacceptable, given the public health risk associated with false underdiagnosis and potential community transmission by the missed patients.

One caveat bears special mention. The “cough” variable in the Taiwanese rule is time-dependent. Specifically, it does not register a score of −2 if cough developed after the onset of fever, that is, only those who had a cough before becoming febrile or when they became febrile were counted. Therefore, as a sensitivity analysis, we randomly reassigned (through 100 iterations) those in the Hong Kong sample who reported the symptom of “cough” to “no cough,” by varying degrees. In our sample, 54.7% were originally recorded as having a cough. To test for the robustness of our findings by taking this caveat in symptom definition into account, we lowered this proportion to 0.1 to 0.4 by simulation, as described. Although sensitivity improved when progressively fewer patients were recorded as having a cough, specificity was lowered correspondingly and the area under the receiver-operating characteristic curve did not change appreciably. Thus, it is unlikely that such variation in the definition of cough is solely responsible for the lack of generalizability of the Taiwanese decision rule.

In addition, Ma and colleagues commented that some of our patients might have acquired SARS after hospital admission. This is possible but unlikely. Almost all patients had a strong contact history and radiographic changes before admission, but many did not have a temperature of 38 °C or greater at admission.

In sum, we conclude that our clinical prediction rule is superior and has excellent external validity. Therefore, it can and should be adopted as the standard to aid health authorities worldwide in planning for a possible return of SARS, especially in regions without adequate access to laboratory facilities or rapid diagnostic testing.