Principles of Signal Detection and Data Mining and their Application to Psychotropic Drugs

The WHO database of drug adverse reactions has over two million case reports of suspected adverse drug reactions, and is the largest of its kind containing spontaneous reports received from the national centres of 73 countries. A technique known as the Bayesian Confidence Propagation Neural Network (BCPNN) has been in routine use for data mining the WHO database as part of the signal detection process since 1998. Data mining is used to enhance the detection of previously unknown possible drug-ADR relationships, by highlighting combinations which stand out quantitatively for clinical review. The use of Bayesian statistics facilitates effective signal detection in this type of data. The Uppsala Monitoring Centre maintains and analyses the WHO database on behalf of the WHO Programme for International Drug Monitoring, and circulates signals in a confidential document to the national centres participating in the program. A possible relationship between antipsychotic drugs and myocarditis was highlighted with the BCPNN in 2001, and other examples of signals circulated through the program include olanzapine with granulocytopenia and venlafaxine and extrapyamidal disorder.

The properties of the neural network allow more complex patterns in the data to be discovered without preconceptions. To demonstrate the usefulness of the method haloperidol ADR reporting was examined and several patterns were detected, including a pattern of symptoms of parkinsonism and the neuroleptic malignant syndrome.

The BCPNN has been used to enhance rather than replace traditional signal detection practices used on the WHO database. Signal detection in spontaneous reporting is a hypothesis generating technique and cannot be used for hypothesis testing. Therefore detected signals are possible relationships that should be investigated further using other techniques.