The data for the student debt was obtained from the US Department of Education database (IPEDS), where the data is available as the average amount of loan aid received by the typical borrower for that year.
In addition, the response rate for year 2 students was lower than anticipated, and in the analysis there were some missing data for the student characteristics variables.
Cases were created for their system using real examinations notes, history notes, test results, and X-rays by expert clinicians, who also provided feedback and meta-data for the students who use the system.
They drew upon the final results data for these students and the other half of the clinical school cohort, which was not included in the pilot project, to form a comparative view of students' academic performance.
Demographic data for the enrolled students are shown in Table 1.
For students' individual narratives, interview transcripts were used as one set of data; for the new student group, this was supplemented by sections in observations where they had told personal stories.
The improvement using the 110-h subset is modest but when the amount of untranscribed data for training the student DNN is increased to 300 h, soft target training is more effective and the WER is reduced by 7.5 and 9.6%% relative.
Looking at the data for the whole student population, we observed a trend whereby three clusters of scores became apparent.
The complete 3D-PAT data set for the student study contained 28 and 66 successful validity and reliability trials, respectively.
We note that, in exploratory factor analysis by principal axis factoring and oblique rotation for our data, four factors emerged, accounting for 41.8% of variance explained, and that these factors were effectively one-half of each of the two domains of the data for college students from the USA (data not shown) [ 12, 16, 17].
