Natural Speech Recognition at the Kaiser Franz Josef Hospital, Vienna

The reporting process
Producing detailed reports for referring clinical departments and health insurance providers prior to starting therapy is an important time and cost factor. The radiology department produces around 75,000 such reports a year. Traditionally, they had to be dictated, handed over to a secretary, typed, proof-read, corrected, printed and finally signed by the reporting doctor - a procedure which involved much time and effort and a prime candidate for simplification and an increase in efficiency.

The department started to look for an efficient, easy-to-use software solution which could be effectively integrated with existing work flows and procedures. As Wolfgang Kumpan, Head of the Central Radiology Institute at KFJ stresses, "In order to achieve user acceptance, it was crucial that the introduction of the new technology would not lead to an increase of work for any of the occupational groups (radiologists, secretaries) involved in the reporting process".

Another objective was to get all urgent reports completed by the radiologists themselves during night duty or on weekends when there is no secretarial support. Treating acute patients at weekends meant producing hand-written (often difficult to read) provisional reports which would get typed up the following week.

The solution
The hospital chose Philips Speech Processing because it was the only supplier at that time who offered a solution which allowed the radiologists to produce reports either on their own or with secretarial support. Additional criteria were the high recognition rate, user-friendliness and ease of integration with the hospital's existing IT infrastructure. As a result, 22 workstations with the Philips SpeechMagic" powered speech recognition system were integrated in existing IT systems provided by Siemens and the Vienna Hospital Association.

Radiologists now dictate reports into a special microphone - the SpeechMike - with an integrated trackball and mouse buttons connected to a Windows workstation. The speech recognition server converts the resulting sound file into text with a very high recognition rate, achieved with the help of an background lexicon containing around 300.000 words, and a customised component with 64.000 entries containing important radiological terms. The speech recognition system has individual acoustic reference files for each user which are constantly updated, and the text generated by the system can be either proof-read and formatted by the radiologist or sent electronically to a secretary for correction. After further validation, a final report in electronic form is made available to the reporting physician.

Implementation
First tests were carried out in February 1998 with a target to produce one third of all reports with the system by the end of 1998. In fact, that target was reached in April and by the end of 1998, two thirds of the reports were produced via speech recognition. After a short period of training and customisation the recognition rate was around 95% with 23 radiologists and 5 secretaries. "It took only about six months that the reporting with speech recognition became total routine", says Wolfgang Kumpan. Now 98% of all reports are produced via speech recognition. "This application is so useful because the recognition rate - which is above 90 per cent - is never critical, as after proof-reading it is always 100%", observes Walter Forsthuber, Director of Philips Speech Processing. The sound file of the dictation is attached to the text file, and they can be read and listened to simultaneously.

Benefits
The introduction of speech recognition there has made audio tapes obsolete and has eliminated illegible referrals. "Tests showed that secretaries are 40% more productive with the help of speech recognition compared to the traditional typing from dictation", says Wolfgang Kumpan. The average production turnaround time for reports decreased from 13 to 8 hours and in many cases the report is ready before the patient returns to the clinical department. "Due to this the therapeutic reaction time and the hospital stay of the patients is reduced", says Wolfgang Kumpan.

Both buyer and technology vendor profited from the co-operation. Although the speech recognition system had already product status, Philips could further improve it due to the knowledge gained during the the project, "We improved the installation and adaptation of lexica and acoustic references in particular", says Walter Forsthuber. "The improvements generally brought about a simplification and a more user-friendly structuring of the processes, mainly through a large degree of automation."

Recommendations
Walter Forsthuber recommends other companies who consider the introduction of a HLT application"to concentrate on the introduction of the new product; it is not good to deal with it as a sideline." And Wolfgang Kumpan says "The introduction of the new technology should be fluid and gentle and not cause more work for any of the occupational groups involved."

Downloads of the full text of the success story in Word [35K] or PDF [16K] formats are available.