A Stratified Model for Psychosis Prediction in Clinical Practice

Abstract

Objective: Impaired cognition is an important dimension in psychosis and its at-risk states. Research on the value of impaired cognition for psychosis prediction in at-risk samples, however, mainly relies on study-specific sample means of neurocognitive tests, which unlike widely available general test norms are difficult to translate into clinical practice. The aim of this study was to explore the combined predictive value of at-risk criteria and neurocognitive deficits according to test norms with a risk stratification approach. Method: Potential predictors of psychosis (neurocognitive deficits and at-risk criteria) over 24 months were investigated in 97 at-risk patients. Results: The final prediction model included (1) at-risk criteria (attenuated psychotic symptoms plus subjective cognitive disturbances) and (2) a processing speed deficit (digit symbol test). The model was stratified into 4 risk classes with hazard rates between 0.0 (both predictors absent) and 1.29 (both predictors present). Conclusions: The combination of a processing speed deficit and at-risk criteria provides an optimized stratified risk assessment. Based on neurocognitive test norms, the validity of our proposed 3 risk classes could easily be examined in independent at-risk samples and, pending positive validation results, our approach could easily be applied in clinical practice in the future.

Introduction

Neurocognitive disturbances are regarded as a core component of psychosis;^1,2 with a recent meta-analysis reporting global cognitive impairments in schizophrenia patients being consistently present in studies over decades and around the world.³ Neurocognitive deficits mainly develop and intensify in the prodrome and early years following diagnosis before settling into a stable pattern of pronounced deficit.⁴ Thus, the first years of psychosis are critical—with regard to cognitive decline and also with regard to many other domains that show a similar pattern, such as psychosocial and occupational functioning.⁵ Yet, psychoses often remain untreated for extended periods.⁶ As the duration of untreated psychosis (DUP) or illness (DUI; including the prodrome) is associated with worse functioning, more symptoms, cognitive impairments, and lower quality of life,^7–10 increasing efforts have been undertaken to shorten DUP and DUI by early detection and intervention. Ideally, early detection and intervention would prevent the onset of frank psychosis and would be early and efficient enough to counteract the cognitive and psychosocial deficits that develop during the prodromal phase.^11,12

Patients symptomatically at risk of psychosis exhibit lower neurocognitive test performance in several domains including general intelligence, processing speed, attention, executive function, and memory.^12–16 At the mean group level, these deficits are mostly intermediate between the performance of healthy individuals and those diagnosed with schizophrenia.^12–14 Furthermore, they are more pronounced in patients considered in a late risk stage (mainly by attenuated psychotic symptoms) than in those considered in an early stage (predominately by subjectively reported cognitive and perceptive basic symptoms).^17,18 In converters, baseline performance is even lower than in nonconverters.^13–16 Thus, neurocognitive baseline deficits are promising candidates for an estimation of risk of conversion. Furthermore, even beyond conversion, cognitive functioning is an important factor for understanding and predicting functional status in at-risk patients¹⁹ and patients with the full-blown disorder.²⁰

“Impaired cognition” was therefore introduced as one of 8 dimensions of psychosis symptom severity proposed in Section III (ie, “Emerging Measures and Models”) of the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5).^21,22

Therein, severity of impaired cognition is defined on a 5-point Likert scale (scores between 0 and 4) using SDs from the age-adjusted mean on neurocognitive tests, with 1 SD or more below mean indicating moderate-to-severe impairment (see supplementary table S1).²¹ Generally, in neurocognitive and other psychological tests, the normal range is defined as ±1 SD from the mean and encompasses 68.26% of people if test scores are normally distributed in the population (see supplementary figure S1). The use of normative scales allows the clinician to determine if an individual differs from the majority of the reference population, defined, eg, by age, gender, education, or socioeconomic status. Such norms are readily available for most neurocognitive tests commonly used in psychosis research. To date, however, studies have predominately reported mean performances, which were calculated based on the raw scores of their sample. These group means generally differ between studies, sometimes considerably. Furthermore, neurocognitive group differences, eg, between converters and nonconverters, can become significant even if both group means still lie within the normal range of the general population (see supplementary figure S2). Consequently, such data provide little information on the presence of actual neurocognitive performance deficits according to norms. Yet, clinicians need this information to be able to classify individual test performances in their daily practice, eg, into estimates of risk for conversion. Despite the potential benefit that the use of test norms would provide for clinical practice, cognitive deficits based on such norms have been understudied in psychosis research. One exception is a study on at-risk patients that explored potential neurocognitive predictors not of conversion to psychosis, but of functional outcome.²³ In addition to group mean analysis, this study also defined deficits in verbal performance, IQ, and verbal memory by standardized IQ scores ≤ 85. Deficient performance was found in 57.9% of the poor outcome group, but only in 26.5% of the good outcome group.²³

Reviews and meta-analyses of other prediction studies based on group means generally reported heterogeneous findings with regard to neurocognitive predictors of conversion.¹² The most promising predictors, however, seem to be a lower performance in processing speed, verbal fluency, visual and verbal memory, and working memory.^{12–16,19,23–25} These studies, however, rarely considered potentially important interactions between at-risk criteria and neurocognitive performance. Only 2 studies so far have studied both psychopathological and neurocognitive predictors.^24,25 Although both studies differed considerably in the selection of predictors and, consequently, in the generated prediction models, each reported a benefit from combining neurocognitive and psychopathological predictors. However, rather than using at-risk criteria and norm-related neurocognitive deficits, which would allow individual risk classification, sample-dependent subscale mean scores of negative or positive symptoms and neurocognitive tests scores were used as predictors.^24,25

Therefore, our aim was to develop an easy-to-apply risk stratification model^26,27 based on 24-month follow-up data of patients symptomatically at risk for psychosis for routine clinical use. Such a model should allow an estimation of conversion risk for individual patients based on the combination of baseline neurocognitive deficits (defined as more than 1 SD according to test norms) and single or combined at-risk criteria.

Method

Sample

The sample consisted of 97 primarily adult patients aged between 16 and 40 years seeking help for mental problems at the Cologne Early Recognition and Intervention Centre for mental crisis (FETZ). All patients fulfilled either ultrahigh-risk (UHR) criteria including (1) “attenuated psychotic symptoms” (APS), (2) “brief limited intermittent psychotic symptoms” (BLIPS), and (3) a “state-trait criterion” according to the Structured Interview for Prodromal Symptoms (SIPS)²⁸ or the basic symptom criterion “cognitive disturbances” (COGDIS) according to the Schizophrenia Proneness Instrument, Adult version (SPI-A).²⁹ The SPI-A, SIPS, and a battery of neurocognitive tests (see below) are routinely administered as part of the clinical diagnostic protocol of the FETZ;³⁰ raters were initially trained (concordance rate with expert rating after 10 training sessions was 91%) and supervised by Frauke Schultze-Lutter in the SPI-A and SIPS assessments. All patients provided informed written consent for participation in the study.

Exclusion criteria for all patients were current or past diagnosis of any psychotic disorder according to DSM-IV criteria; diagnosis of delirium, dementia, amnestic or other cognitive disorder, mental retardation, mental disorders due to a general medical condition or substance-related disorder according to DSM-IV; alcohol or substance use disorders within the past 3 months according to DSM-IV (assessed with the Structured Clinical Interview of DSM-IV (SCID),³¹ no toxicology screen conducted); and diseases of the central nervous system (inflammatory, traumatic, epileptic). Patients were monitored for a conversion to psychosis according to the SIPS over a mean duration of 15.80 months (SD = 8.37; range = 1–37). In case of a conversion, the type of psychotic disorder was assessed with the psychosis section of the SCID.³¹ Fifty-three (54.6%) patients did not convert to psychosis over a mean time of 19.96 months (SD = 6.18; range = 3–37). Forty-four patients (45.4%) converted to a psychotic disorder according to DSM-IV criteria within 10.80 months on average (SD = 7.96; range = 1–24): 36 (81.8%) developed schizophrenia, 2 (4.5%) delusional disorder, 1 (2.3%) schizophreniform disorder, 1 (2.3%) psychosis not otherwise specified, 1 (2.3%) schizoaffective disorder, 1 (4.5%) major depressive disorder with psychotic features, and 1 (2.3%) bipolar disorder with psychotic features. Altogether, the majority of patients (86.4%) developed a schizophrenia spectrum disorder (n = 38) and only 6.8% developed an affective (n = 3) or other psychosis (n = 3). Converters and nonconverters did not differ at baseline in sociodemographic characteristics or the presence of any nonpsychotic DSM-IV axis I disorder although converters had more frequently reported recurrent brief depressive disorder (table 1; for more details on distribution of axis-I disorders see supplementary table S2). Furthermore, the frequency of deficits in estimated premorbid IQ did not differ between converters and nonconverters although a small-to-moderate effect of group on the mean estimated premorbid IQ was present in favor of nonconverters (Rosenthal’s r = .242; table 1). At baseline, all participants had never been treated with an antipsychotic. During the follow-up period, however, converters more frequently received antipsychotic medication (this includes medication for frank psychosis after conversion) and, expectedly, still more frequently took antipsychotic medication at follow-up. The rate of psychotherapy did not differ between converters and nonconverters (for treatment details see supplementary table S2).

Table 1.

Characteristics of Sample

	Converters (n = 44)	Nonconverters (n = 53)	P a
Age (y)
Mean (±SD)	24.1 (±5.7)	25.3 (±5.3)	0.200
Median (range)	22.8 (16.3-39.3)	24.7 (17.1-37.1)
Gender, % male	65.9	64.2	0.857
Partnership, %
Single	93.2	92.5	0.459
Married/steady partner	4.5	7.5
Separated	2.3	0
Graduationb, %
None	7.0	0	0.086
Certificate of secondary education (10 y)	9.3	3.8
O-level (10 y)	20.9	13.2
Vocational baccalaureate diploma (12 y)	7.0	13.2
A-level (13 y)	37.2	58.5
Still in school	18.6	11.3
Vocational education, %
None	25.0	9.4	0.157
Apprenticeship or similar	11.4	22.6
Master craftsman or similar	2.3	0
College of higher education	2.3	0
University	6.8	7.5
Still in school/training	52.3	60.4
Current occupation, %
No work/education	22.7	17.6	0.537
Regular occupation including education	77.3	82.4
Any current, nonpsychotic DSM-IV axis-I disorderc, %	65.9	59.6	0.526
Premorbid IQ by MWTd
Mean (±SD)	27.8 (±4.5)	29.9 (±3.6)	0.018
Median (range)	28.0 (15-35)	30.0 (20-36)
Deficit according to norms	1	0	0.274

Notes: MWT, German version of the Multiple Choice Vocabulary Test. ³²

^a U-test and 2xk-χ test, respectively.

^bTranslated into British graduations (minimum years of school education required to receive the respective graduation).

^cAs assessed with the German version of the Structured Clinical Interview of DSM-IV axis I disorders. ³¹

^dMWT is a measure of verbal IQ highly correlated with total IQ; MWT values of 6-20 correspond to IQ values of 73–90, of 21–30 to 91–109, of 31–33 to 110–127 and of 34–37 to ≥128.

Assessments

The neurocognitive test battery, previously described in detail,³³ examines the following domains:

• 1. Attention: Continuous Performance Test, identical pairs version (CPT-IP)^34,35

• 2. Memory: verbal memory with the Auditory Verbal Learning Test (AVLT);^36,37 visual memory with the Rey-Ostrieth Complex Figure Test (ROFT);^38,39 spatial working memory with the Subject Ordered Pointing Task (SOPT)^39,40

• 3. Executive functions: Wisconsin Card Sorting Test (WCST);⁴¹ verbal executive functions with the verbal fluency task⁴²

• 4. Processing speed: Digit Symbol Test (DST)⁴³ and Trail-Making Test (TMT) A and B^39,44

Neurocognitive deficits were defined relative to the normative data provided for each test as (1) more than 1 SD below the mean, (2) a T-score below 40, or (3) a percentile below 16. We note that a T-score below 40 and a score below the 16th percentile both correspond to 1 SD below the mean (see supplementary figure S1). According to these criteria, a neurocognitive deficit in our study equals at minimum a “moderate cognitive deficit” according to the DSM-5’s new dimensional severity assessment; ie, “a clear reduction in cognitive function below expected for age and socioeconomic status.”^21(p744) Details of the test parameters and definitions of the thirteen neurocognitive deficits are provided in table 2.

Table 2.

Neuropsychological Measures and Definition of ‘Neurocognitive Deficits’

Neuropsychological Domain	Test Description	Definition of Neurocognitive Deficit; Adjustment of Norms
Attention	The Continuous Performance Test (identical pairs version; CPT-IP) 34 provided a measure of sustained attention. The signal detection parameter d' was calculated across 300 trials.	CPT-IP—d' (more than 1 SD below mean); norms unadjusted for age. 35
Memory
Verbal memory	The Auditory Verbal Learning Test (AVLT) 36 provided a verbal memory measure for immediate and delayed recall after one to 5 learning trials of word lists.	AVLT immediate recall—no. correct after 1st trial (PR < 16); norms adjusted for age. 37 AVLT trials 1–5—sum no. correct (PR < 16); norms adjusted for age. 37 AVLT delayed recall—no. correct after 30min (PR < 16); norms adjusted for age. 37
Visual memory	A measure of visual memory was provided by the Rey-Osterrieth Complex Figure Test (ROFT). 38 The delayed recall performance was scored according to L.B. Taylor’s criteria.	ROFT—delayed recall (more than 1 SD below mean); norms adjusted for age. 39(p827)
Working memory
Spatial working memory	During each trial of a computerized version of the Subject Ordered Pointing Task (SOPT) 40 subjects had to point to 1 of 12 objects, and the relative positions of the objects varied randomly across trials. Across 3 sessions of 12 trials the number of errors (pointing to an object already chosen on a previous trial) was calculated.	SOPT—no. errors (more than 1 SD below mean); norms adjusted for age. 39(p475)
Executive functions
Set shifting and problem solving	The percentage of perseverative and nonperseverative errors made in the Wisconsin Card Sorting Test (WCST) 41 provided a measure of set shifting and problem solving.	WCST—% errors (T < 40); norms adjusted for age. 41
Verbal executive functions	Verbal executive functions were measured by a verbal fluency task (a lexical and a semantic category task). 42 For lexical verbal fluency as many words as possible beginning with a S and for semantic verbal fluency as many animal names as possible had to be produced within 1min.	Lexical verbal fluency—no. correct (PR < 16); norms adjusted for age. 42 Semantic verbal fluency—no. correct (PR < 16); norms adjusted for age. 42
Processing speed	The digit symbol test (DST) 43 and trail-making test (TMT) A and B 44 provided measures for the speed of visual information-processing and visuomotor coordination.	DST—no. correct (T < 40); norms adjusted for age. 43 TMT A and B—time in s (more than 1 SD below mean); norms adjusted for age. 39(p659) Ratio TMT B/A (more than 1 SD below mean); norms adjusted for age. 39(p663)

Notes: PR, percentile; M, mean. All neuropsychological measures were already described in an earlier publication of the same sample.^33(p44–45)

In addition to the neurocognitive deficit predictors, the following predictors of at-risk criteria were considered: “any APS,” “only APS,” “any COGDIS,” “only COGDIS,” and the combinations “APS + COGDIS” and “BLIPS + APS + COGDIS.” BLIPS was never observed without both APS and COGDIS. The state-trait criterion of the ultrahigh-risk criteria was not considered as a potential predictor because it did not appear in our sample.

Data Analyses

Follow-up periods were censored at the end of month 24 to establish a defined reference period of risk estimation across all subjects. Using SPSS 20.0, potential binary predictors for the final Cox regression were selected in several steps in order to derive a parsimonious model^26,45: first, predictors were computed individually in Cox regression and selected for further analyses when changes of the −2 log-likelihood of the model and the Wald statistic became significant at a liberal level (P < .15). Next, forward multivariate Cox regression analyses were performed within each subgroup of predictors (at-risk criteria and each cognitive domain) at a significance level of P < .10 for further selection of predictors. The assumption of proportionality of the hazard function over time was tested prior to each Cox regression analysis and maintained for all potential predictors. The remaining predictors were analyzed together forward and backward to exclude effects of blocking (P < .05), restricting the maximum number of predictors entering the final model to a 1:5 ratio of number of predictors to events, ie, to 8 predictors.⁴⁶ The internal validity of the final prediction model was assessed with the bootstrap resampling technique.^47,48

Following the suggested approach for risk stratification,^26,27 the resulting Cox equation was used to calculate individual prognostic scores. After Kolmogorov-Smirnov testing for normal distribution (P < .001), a 2-step cluster analysis was conducted for the nonnormally distributed prognostic scores to identify the number and types of profiles across the sample that could be interpreted as risk classes. The final Cox regression model with and without MWT-B deficit was compared to assess the additional influence of premorbid IQ on predictor selection. Furthermore, it was tested whether the inclusion of premorbid IQ deficit led to different risk classes in the 2-step cluster analysis.

Risk classes were subsequently compared for their conversion rate and time to conversion by Kaplan-Meier analysis (P < .05).

Results

After the 2 selection steps (see supplemental table S3), “APS + COGDIS,” “AVLT immediate recall,” “semantic verbal fluency,” and “DST deficit” remained for further Cox regression analysis. Of these, “APS + COGDIS” and “DST deficit” were selected for the final model (table 3). The omnibus test confirmed that the model as a whole was highly significant (−2LL = 353.686, χ²(2) = 10.473, P = .005). The resulting equation was [(−1.582) × “APS + COGDIS”] + [(−0.642) × “DST deficit”]. The bootstrap approximation with 5000 iterations yielded a bias-corrected and accelerated confidence interval, which demonstrated the generated model’s robustness because zero is not between the lower and upper bound (“APS + COGDIS”: 95% CI: (−13.357)|(−0.686), P = .028; DST deficit”: 95% CI: (−1.291)|(−0.069), P = .031).⁴⁸

Table 3.

Final Cox Regression Model

		Beta	SE	Wald (df = 1)	P	HR	95% CI
At-risk criteria	APS + COGDIS	−1.582	0.725	4.764	.029*	0.206	0.050–0.851
Memory	AVLT immediate recall	—	—	—	.110	—	—
Executive functions	Semantic verbal fluency	—	—	—	.233	—	—
Processing speed	DST	−0.642	0.314	4.186	.041*	0.526	0.285–0.973

Notes: HR, hazard ratio; APS, attenuated psychotic symptoms; COGDIS, cognitive disturbances; AVLT, Auditory Verbal Learning task; DST, digit symbol test. Cox regression, method “Wald forward” and “Wald backward”, all predictors were compared at once and selected if P ≤ .05 (selected binary predictors are in bold type).

*P ≤ .05

The 2-step cluster analysis of the Cox regression’s prognostic scores revealed 4 clusters. The Silhouette measure of cohesion and separation was 1.0, which indicates an excellent overall goodness-of-fit. The ratio of the largest to the smallest cluster was 5.38 (cluster 2, n = 43, 44.3%; cluster 4, n = 8, 8.2%). These 4 clusters translated to the following predictor constellations: (I) neither “DST deficit” nor “APS + COGDIS” (n = 9; 0 converters), (II) only “DST deficit” (n = 8; 2 converters), (III) only “APS + COGDIS” (n = 37; 17 converters), and (IV) both “DST deficit” and “APS + COGDIS” (n = 43; 25 converters).

These results did not change when the covariate “MWT-B deficit” was forced into the Cox model: “APS + COGDIS” and “DST deficit” were still selected into the final model, and exactly the same 4 risk classes were revealed in the subsequent 2-step cluster analysis.

Subsequent Kaplan-Meier analysis revealed significantly different survival curves for the risk classes II–IV (Tarone-Ware test: χ²(3) = 10.784, P = .013); class I did not enter this analysis due to the lack of any converters. The mean time to conversion in class II was 19.0 months (95% CI: 13.0–25.0), 18.8 months in class III (95% CI: 16.1–21.5), and only 15.4 months (95% CI: 12.7–18.1) in class IV. The cumulative hazard rates for conversion were 0.0 in class I, 0.29 in class II, 0.75 in class III, and 1.29 in class IV; the hazard function is shown in figure 1.

Fig. 1.

Hazard function of the 3 risk classes by a hierarchical cluster analysis. Follow-up periods exceeding 24 months were considered censored at the end of month 24. APS, attenuated psychotic symptoms; COGDIS, cognitive disturbances; DST, digit symbol test.

Discussion

An earlier detection of and, consequently, intervention in emerging and early psychosis is crucial to the aim of reducing the immense overall burden still associated with this group of disorders.¹¹ Ideally, an early intervention would not only help to prevent full-blown psychosis but also prevent or ameliorate the psychosocial and occupational functional decline that precedes psychosis onset. To this end, both sensitive and specific risk criteria have to be developed that should be transferable to clinical practice. And while current symptomatic at-risk criteria according to the UHR and basic symptom approach have proven to be good starting points, studies so far clearly indicate that additional predictors are needed to increase specificity.¹²

With this aim and in hopes of advancing earlier studies, we investigated whether combining at-risk criteria and the presence of neurocognitive deficits according to the respective test norms improved psychosis prediction. This approach defines neurocognitive deficits according to norms, in line with the dimensional severity assessment of clinical phenomena in psychosis included in Section III of the DSM-V.^21,22 Relying on deficits according to the test norms provided for most neurocognitive tests rather than variable group means from different study samples, this approach translates easily into clinical practice.^21,22 Thus, our study is an important extension to the existing body of neurocognitive prediction studies^19,23–25 and the available approaches to individualized risk estimation.^26,27

Altogether, 2 predictors were selected for the final, internally validated, parsimonious Cox regression model: (1) at-risk criteria in terms of the concomitance of APS criteria according to the SIPS and cognitive disturbances (COGDIS) according to the SPI-A (APS + COGDIS) and (2) a processing speed deficit according to the DST (a simple and quick-to-apply paper-pencil test; see supplemental figure S3). Hence, the risk of psychosis was lowest when neither a processing speed deficit nor the 2 at-risk criteria were present, and highest and most immediate when both were present; taken individually, the sole presence of a processing speed deficit indicated a lower risk than the sole presence of “APS + COGDIS”.

Regarding neurocognition, the selection of a processing speed deficit according to the DST into the prediction model is in line with recent data. Impairment in processing speed has been shown to distinguish well between healthy controls and help-seeking at-risk patients.^{13,15,17,24,49} Further, an impairment in processing speed was most pronounced in a community-based sample of non-help-seeking adolescents who met ultrahigh-risk criteria (mainly APS) according to the SIPS;⁵⁰ and in a large birth cohort study, impairment in processing speed at age 8 was a particularly strong predictor of psychotic experiences at age 12.⁵¹ Impairment in processing speed has also been reported as the earliest emerging neuropsychological deficit in emerging psychosis;⁵² and, after the first manifestation of frank psychosis, schizophrenia patients are most impaired in processing speed.^3,52,53 Therefore, processing speed has repeatedly been suggested as a core neurocognitive deficit that might mediate other cognitive deficits in schizophrenia.^50,52,54,55 Additionally, in at-risk samples, processing speed deficits correlated strongly with self-perceived impairments in stress tolerance cross-sectionally,³³ and predicted poor social and role functioning longitudinally.^23,49,56 Therefore, this simple neurocognitive measure seems to provide a robust signal of impairment related not only to the presence of psychosis^3,52,53 and poor functional outcomes^23,49,56 but also to the risk of psychosis.^{13,15,17,24,49}

With regard to the assumed central role of processing speed deficits,^50,51,54,55 its neural basis also makes it a likely candidate for a core deficit of psychosis. Contrary to many other neurocognitive functions, processing speed is not thought to be strongly connected to a particular brain region (as is the case for memory, which is strongly associated with the hippocampus formation) but rather related to the general neuroconnectivity of the brain. Thus, processing speed is thought to reflect a process of integration and coordination between distributed networks.⁵⁴ Indeed, in a tractography imaging study on healthy controls and brain-injured patients, it was closely related to the structural integrity of major white matter tracts.⁵⁷ In line with findings in psychosis⁵⁸ and the disconnection hypothesis of schizophrenia,⁵⁹ disrupted white matter integrity has also been reported in studies of UHR samples.⁶⁰ Further evidence of the link between cognition and cortical structure in UHR patients comes from a neuroimaging study that found that the decline in neurocognitive measures (incl. processing speed measures) is associated with alterations in interconnecting white matter.⁶¹ Thus, processing speed deficits may signal the cortical disconnectivity that plays a role in the development of psychosis.

When translating our findings for clinical practice, it is essential to consider that the predictive value of processing speed deficits depends on the test used.¹³ In our study, processing speed as assessed by the DST was selected as a predictor; processing speed as assessed by the TMT was excluded in the first selection step (see supplemental table S3). Similarly, studies that utilized a symbol coding task as a measure of processing speed generally showed the most pronounced differences between at-risk and healthy controls.^{13,15,17,24,49}

On the psychopathological level, selecting the “APS + COGDIS” combination into the prediction model fits well with earlier results indicating an improvement in psychosis prediction by combining these 2 at-risk criteria.^12,26 Our study further highlights the importance of subjective cognitive disturbances independent of the objective cognitive disturbances as these 2 measure do not correlate.³³ Furthermore, the fact that the exclusive presence of the “APS + COGDIS” combination is more specific for a later conversion than the exclusive presence of a processing speed deficit highlights the importance of psychopathology in psychosis prediction.

Strengths and Limitations

Beyond the potential of our results to easily translate into practice, a strength of our study is the high rate of critical events—conversions—in this sample. The distribution of affective and nonaffective psychoses in converters is in line with findings of a recent meta-analysis reporting 73% schizophrenia spectrum disorders, 11% affective psychoses, and 16% other psychoses for converters in UHR studies with a higher rate of schizophrenia spectrum disorders in studies also using the basic symptoms.⁶² The high percentage of conversions, which lies within the upper range of conversion rates reported from other studies,¹² allowed us to investigate the impact of several predictors from a neurocognitive test battery covering almost all domains reported in other studies on neurocognition in at-risk samples. This avoided bias toward the selection of certain neurocognitive predictors.

One limitation of our study was the duration of the observation time. The true outcome of presumed nonconverters beyond the 24-month follow-up period remains unclear. Thus, future studies will need to test the reliability of our model in other samples, as well as its validity within longer time frames. However, conversion studies with longer follow-up periods have consistently reported that the majority of conversions occur within the first 2 years after baseline.¹²

Another potential limitation is the difference in the quality of available test norms; some of which—such as those of SOPT—have to be regarded as poor. However, norms provided for the tests of the Wechsler Adult Intelligence Scale (WAIS) incl. the DST are excellent and possess good age-adjusted standardization properties.⁴³ Another advantage of the WAIS is its adaption to many languages with high-quality norms being provided for a number of countries, including Germany.

A possible clinical, yet unlikely, limitation is the potential influence of cannabis use on neurocognitive test performance.⁶³ In line with other UHR studies using the SIPS,^26,30,64,65 cannabis use at baseline was infrequent in our study, reported by only 2 persons. However, in cannabis-using patients, clinicians should be aware that cannabis use may impact test performance.

The reported association of poor performance in a variety of neurocognitive tests, including the DST, and the diagnosis of recurrent brief depression⁶⁶ might convey another limitation to our results because recurrent brief depression was significantly more frequent in converters at baseline than in nonconverters. Yet, poor performance in the DST in particular was related to a previous history of major depression in patients with recurrent brief depression.⁶⁶ Patients with such a history, however, were all nonconverters in our study. Thus, a negative rather than a positive mediating effect of the presence of recurrent brief depression on the psychosis-predictive value of the DST in our sample has to be assumed.

Conclusions

In summary, our findings advance previous studies of the early detection of psychosis in 2 ways: first, our approach of using neurocognitive deficits based on norms rather than sample-dependent group means is unique. Moreover, our model, which is based on 2 well-established at-risk criteria predictors and a neurocognitive predictor known to be a core neurocognitive deficit in both at-risk and psychotic patients, could easily be validated in other existing at-risk samples using the readily available test norms.

Second and most importantly, pending such validations in independent samples, it might allow an estimation of current risk for psychosis in the individual patient in clinical practice with no need for specific, expensive techniques with limited availability.

Supplementary Material

Supplementary material is available at http://schizophreniabulletin.oxfordjournals.org.

Funding

German Research Foundation (KL970/3-1,2 to J.K. and F.S.L.).