Purpose

   To assess how three diagnostic logics perform in terms of sensitivity, specificity and avoided harm.

Methods

   The theoretical validity of three diagnostic approaches was compared: i) an algorithmic strategy following either a linear or a branched pathway; ii) a serial threshold-based strategy and iii) a parallel threshold-based strategy. The diagnostic management of immune compromised HIV patients presenting with headache as chief complaint in a developing country was used as pivot problem. Possible diagnoses were limited to five dangerous but treatable diseases, including cryptococcal meningitis, cerebral toxoplasmosis, tuberculous meningitis, bacterial meningitis and malaria. Only key clinical characteristics available in a low or moderate income country were considered.

   Performance of the strategies was estimated based on prevalence, sensitivity of disease characteristics, mortality and morbidity rates of disease and treatment, treatment effectiveness, and relative weight -as perceived by clinicians- of false negative and false positive diagnoses. Data was based on literature retrieval and Delphi surveys. Performance of each strategy was measured in terms of sensitivity, specificity and harm related to mortality and morbidity. Cost was not considered as care was free of charge. Baseline and sensitivity analyses as well as a Monte Carlo simulation were performed.

Results 

   Through the baseline analysis the parallel threshold-based approach appeared to be the one that leads to the highest sensitivity and engenders the lowest harm in terms of mortality and morbidity.

	sensitivity	specificity	harm in terms of mortality and morbidity (relative to the lowest)
Linear algorithm	74%	96%	1,44
Branched algorithm	47%	95%	2,41
Serial threshold-based	45%	84%	2,35
Parallel threshold-based	94%	64%	1

These findings were corroborated by both the sensitivity analyses and the Monte Carlo simulation.

Conclusion

   A threshold-based logic is conceived to find the optimal balance between false negatives and false positives in order to minimize harm. This study proves that including this logic in a serial algorithm is not sufficient: all diseases should be analysed in a parallel way. The relative superiority of this parallel threshold based approach is possibly limited to a presenting symptom pointing to life threatening diseases.