Download 1. Introduction

Predicting relapse of schizophrenia
Petr NALEVKAa1
University of Economics, Prague, Czech Republic
a
Abstract. Schizophrenia is a serious mental disorder which tends to follow a
relapsing course leaving a devastating effect on the patients [1]. This article
proposes symbolic data-mining methods to predict early stages of schizophrenia
relapse for timely intervention.
Keywords. relapse of schizophrenia, symbolic data-mining, Youden
1. Introduction
Relapse of schizophrenia is preceded by various prodromal symptoms [2]. Timely
detection of such symptoms may help predict and prevent onset of the relapse [2,3].
This principle is used in the ITAREPS program [4], which collects patients' 2 condition
self-evaluations sent by SMS messages and automatically informs the outpatient
psychiatrist in case patient's worsening exceeds an expert-defined threshold. Since
2006, in two consequent studies (mirror-design and double-blind) ITAREPS gathered a
unique set of demographic and diagnostic data from over 400 patients.
2. Objectives
Employ data-mining methods to improve prediction performance of the ITAREPS
program. Especially, lower the amount of false alerts and search for interesting patterns
to improve knowledge of schizophrenia development in different patient types.
3. Methods
Symbolic data-mining methods are used to discover patterns which are successful
predictors but also interpretable for the psychiatrists. Applicability of such methods has
been proven on similar system failure prediction problems [5,6].
Patients' condition history is split into temporal windows either preceding relapse
(positive) or not (negative). Different patterns are generated and their ability to
distinguish between positive and negative windows is tested using the Youden's index
(Sensitivity + Specificity - 1) [7,8] in order to minimize error. Patterns consist of highlevel categorical features combined using logical connectives. Features describe
1
2
Corresponding author
In addition, evaluation SMS messages, are also sent by patient's carer – a selected family members.
development of patient's condition within windows such as the level, trend or stability
of worsening for individual symptoms or groups of symptoms.
Patterns are evaluated for each patient group which is established based on
demographic and diagnostic data. This helps to fight noise in data, lower errors and
discover knowledge which is specific to different types of patients.
4. Results
In comparison to ITAREPS's Youden = 0.226 and specificity = 0.769 the described
method performed at Youden = 0.324 and specificity = 0.873 in a leave-one-out
prediction performance estimate (3 week temporal windows, positive windows in 2-6
weeks before relapse). The overall number of false alerts has been reduced from 1709
to 955. Table 1 shows an example of 3 best patterns discovered.
Table 1. Most successful patterns by training set Youden
Youden
0.58
0.58
0.51
Pattern
married, no cognitive symptoms, carer sympt.1 (high avg.), all carer symptoms (high avg.)
young, non-university, sympt.10 (high avg.), all carer symptoms (low std.deviation)
young, no disorganization symptoms, carer sympt.10 (high average)
5. Discussion and Conclusion
The presented results meet the outlined objectives. Prediction performance estimates
show an improved Youden with a significant increase in specificity over the current
ITAREPS. False alerts are reduced by 44% with 12% false alerts per SMS.
Moreover, table 1 presents an example of discovered background knowledge.
Patients with absent cognitive and disorganisational symptoms seem better suited for
relapse prediction. In addition evaluation done by carers seem a better predictor.
Acknowledgements. This project was supported by Grant Agency of Czech republic, GACR
201/08/0802.
References
[1] van Os J, Kapur S. Schizophrenia. Lancet (2009), 374: 635-45.
[2] Birchwood M, Smith J, Acmillan MF, Bridget H. Predicting relapse in schizophrenia: the development
and implementation of an early signs monitoring system using patients and families as observers a
preliminary investigation, Psychological Medicine, 1989.
[3] Fitzgerald P. The role of early warning symptoms in the detection of relapse in schizophrenia, Australian
and New Zealand Journal of Psychiatry, 2001.
[4] Španiel F, et al. The Information Technology Aided Relapse Prevention Programme in Schizophrenia: an
extension of a mirror-design follow-up,Int J Clin Pract., 2008.
[5] Weiss G, Hirsch H. Learning to Predict Extremely Rare Events, Rutgers University New Brunswick,
2000.
[6] Vilalta R, Sheng M. Predicting Rare Events In Temporal Domains, IEEE International Conference on
data-mining, 2002.
[7] Pepe MS. Evaluating technologies for classification and prediction in medicine, Statistics in Medicine,
24(24), 3687-3696, 2005 .
[8] Youden WJ. Index for rating diagnostic tests, Cancer (1950), 3: 32-35