EL MARCO DE LA CONTABILIDAD DE GESTIÓN EN LOS CENTROS SANITARIOS
3.6.2. Sistemas de clasificación de pacientes
On average, across each of the conditions, participants made their emotion judge-ments 5.10 seconds after stimulus onset (corresponding to mean stimulus length + 1.93 seconds). Reaction times were not compared across the different feature-attenuation conditions or emotions, since this measure was confounded by differences in stimulus length.
The average confidence rating associated with participants’ judgements, across all of the experimental trials, was 3.68. Rank-based factorial ANOVA revealed a very small but significant effect of Emotion upon confidence ratings, F (4, 820) = 2.61, p = .034, η2= .01. Post-hoc, Bonferroni-adjusted Wilcoxon-Mann-Whitney tests showed that the only significant pairwise difference in confidence ratings was between Sadness and Neutral, with the former being higher (see Table 9). There was no significant effect of Feature attenuation, F (3, 820) = 1.56, p = .194, η2 <.01, nor was there a significant interaction with Emotion, F (12, 820) = 0.90, p = .551, η2 = .01.
Table 9: Pairwise, Bonferroni-adjusted Wilcoxon-Mann-Whitney tests, comparing differences in participants’ confidence ratings, according to the emotion judged.
Anger Fear Happiness Neutral
Overall performance (mean percentage correct) in the pre-test, test and post-test-transfer blocks is summarised for each condition in Figure 23. In general, par-ticipants in each of the different feature processing conditions showed an effect of training with the CI-simulated stimuli, performing better, in all conditions, in the post-test and post-test transfer stimuli blocks, compared to pre-test. In all conditions except for Duration, performance also improved from the post-test to the post-test-transfer phase, implying that participants’ learning was not restricted to the specific stimuli experienced during training. ANOVA confirmed that the effect of test phase upon emotion identification accuracy as significant, F (2, 46) = 20.52, p <.001, η2
= .47. Planned contrasts revealed that the differences between pre-test and the two post-test phases were significantly different (p <.001 for both), whilst the difference between the post-test and post-transfer phases was not significant.
Figure 23: Mean percentage correct emotion recognition, for each condition, across the different testing phases. Error bars show± 1 standard deviation.
Looking at post-test performance, participants did best in the Original condition, as expected, but elsewhere the differences between feature attenuation conditions were minimal. Even though, for example, accuracy in the Duration condition at pre-test was eleven percentage points worse than in the Frequency condition, these differences were largely eliminated by the training phases. By contrast, the training created a noticeable discrepancy between the feature attenuation conditions and the Original condition, with the latter associated with much greater emotion judgement accuracy.
Loglinear analysis, examining the effects of feature processing condition and emotion upon judgement accuracy, produced a final model that retained all effects – likelihood ratio χ2 (0) = 0, p = 1 – indicating that the Processing× Emotion interaction was significant, χ2 (12) = 26.36, p < .010.
Follow-up chi-square analyses revealed significant effects of Emotion upon judgement accuracy in all processing conditions except for Frequency, as summarised in Table 10. This appeared to reflect the fact that the different emotions were recognised with relatively uniform accuracy in the Frequency condition, whereas in the Dura-tion condiDura-tion, for example, Sadness was distinguished with almost four times the accuracy of Happiness.
Table 10: Chi-square tests, examining the influence of stimulus emotion class on emotion recognition accuracy in each of the feature processing conditions.
Condition χχχ222 df p
Duration 21.14 4 < .001*
Frequency 3.16 4 .531
Intensity 13.19 4 .010*
Original 40.34 4 < .001*
*= significant at .050 alpha-level.
To explore the patterns of errors in emotion judgements made by participants, con-fusion matrices were created for each of the feature processing conditions (Figure 24). These show, for each stimulus emotion, the frequency with which participants selected each of the five response options.
Firstly, it is clear that there was a general effect of learning - accuracy for each of the emotions largely increased from pre-test to post-test-transfer. Additionally, the overestimation of Neutral, as observed in Study 1, declined after the training, as indicated by lower ‘false positive’ response rates for Neutral in the post-test-transfer phase compared to the pre-test. There was also some evidence of reduced false positive responding for Anger, although confusions with Happiness increased in all conditions, implying that this particular confusion was highly pervasive and difficult to overcome.
Sadness and Neutral appeared to be the most robust emotions to disruption over-all, with these emotions being relatively well-decoded across all feature attenuation conditions, and also showing quite clear improvement as a result of training. Al-though there was some universal improvement as a result of acclimatisation to the CI-simulated stimuli, there were also specific combinations of feature attenuation conditions and emotions that showed particularly good improvement with training.
For example, recognition accuracy for Sadness stimuli improved by 20.83 and 12.50
percentage points in the Frequency and Intensity conditions, respectively, but by 41.85 in the Duration condition. Similarly, Fear stimuli were better recognised in the Duration condition, and showed more improvement from training compared to the Frequency and Intensity conditions. Conversely, Anger improved to a greater extent and was much more accurately identified in the Frequency and Intensity conditions.
Figure 24: Heat-mapped confusion matrices for speech stimuli during the pre-test and post-test-transfer phases, depicting the percentage of responses in each emo-tion category, for each stimulus emoemo-tion. Columns denote presented emoemo-tions, rows denote emotion judgement responses. Red = higher values, green = lower values.
Lastly, the potential relationships between musical ability (MUSE) and empathy (EQ) measures, and emotion identification accuracy were explored. Controlling for the effect of Feature attenuation condition, multiple linear regression analysis found that none of the MUSE sub-scores, nor the EQ scores, were able to explain a sig-nificant portion of the variance in the percentage of correct emotion judgements at any of the test phases. However, the Pearson correlation between the ‘Index of
Musical Training’ subcomponent of the MUSE and emotion judgement accuracy be-came increasingly strong with training, and reached statistical significance for the post-test-transfer results (Table 11). Specifically, more highly musically-trained indi-viduals appeared to benefit to a greater extent from the training paradigm, tending to register better emotion identification accuracy in the final, post-test-transfer phase.
Table 11: Pearson correlations between MUSE subcomponent ‘Index of Musical Training and overall emotion identification accuracy.
Phase r df p
Pre-test .29 22 .163
Post-test .38 22 .063
Post-test-transfer .42 22 .040*
*= significant at .050 alpha-level.