Q.5. Verificación de la propuesta
Q.5.3. Caso GM2: El IDS en la toma de decisiones estratégica
The rules presented in the previous section enabled the ultrasound reporting system to automatically classify findings as “normal”, “abnormal” or “inconclusive” based only on the words that the sentence contained. When applying these rules on each sentence of the findings in all 100 sample ultrasound reports, the system managed to classify most of the sentences without any problem especially for sentences with nor- mal findings. In order to verify the result of the automatic classification, two medical ultrasound specialists have been approached to seek for their help in manually clas- sifying each finding in the 100 reports so that it can be compared with the automatic classification completed by the system. However, they argued that the classification was not possible thus making it impossible to validate our data.
The specialists required the “normal”, “abnormal” and “inconclusive” classifica- tion to be defined and given certain criteria for them to be able to classify the findings because they argued that the classification of the findings was not binary. We defined “normal” findings as findings that do not need further attention while “abnormal” find- ings was defined as findings that needs further observation. “Inconclusive” findings
on the other hand was defined as findings that could not be determined whether it is normal or abnormal because of reasons such as the organ was obstructed or was not seen and the ultrasound examination might need to be carried out again.
These definitions however were not enough to classify the findings in the reports as the specialists argued that the classification of the findings strongly depends on the context of the referral which includes information such as the patient’s clinical history and the reason for referral. This information is important so that correct interpretation can be made. They provided two examples; (i) a gall bladder polyp could be the cause of symptoms in somebody with pain in that area, but would be an incidental finding in someone asymptomatic and (ii) an endometrial thickness of 6mm would be normal in a 25-year-old but will likely be abnormal in a 70-year-old. Therefore, the absence of the context of referral and other relevant information makes it difficult to classify the findings of the reports. This information was not made available to us because of the privacy and confidentiality of the patient’s data. Therefore, as it stands, the specialists were not able to perform the manual classification because many of the information needed was absent.
The automatic classification was intended to assist the radiologists in completing the structured reports as they will have fewer fields to fill. However, the discussions with the medical ultrasound specialists have made it clear that this automation pro- cess was not possible because of the absence of the necessary information. If manual classification of the findings in the reports could not be completed by the ultrasound specialists, hence, it would not be possible for the system to do the same. As a result, this feature has been abandoned from the system for the time being.
6.6
Chapter Summary
Various ultrasound reporting styles and format have prompted the need to standardise them so that it is better read and understood by the referring clinicians. This chapter presented the development of the medical ultrasound reporting system including the development tools used. It explored the implementation of both RST and AUO in the transformation of the free-form reports to structured form. It first reviewed the 60 re- ports training data which recognised four different types of information that existed in the reports. This has resulted in the identification of the PREPARATION, RESTATE- MENT, JOINT, LIST and ELABORATION relation as the RST rhetorical relations which are relevant to be applied in the transformation process.
Following this, the chapter presented all the steps taken in translating the free-form reports to structured form and how these relations were applied. Next, the chapter discussed the software quality aspect of the overall system and the evaluation of the structured report generator from a pair of specialists regarding the implementation. Their evaluation found the transformation as generally correct. Finally, the chapter discusses an attempt to automatically classify ultrasound findings into “normal”, “ab- normal” and “inconclusive”. However, the data was not validated since the feedback from the ultrasound specialists suggested that this is not possible. This has resulted in the abandonment of the automatic classification process because there were insufficient amount of information for even the specialists to manually classify them.
Conclusion and Future Directions
7.1
Introduction
The final chapter of this thesis summarises the research and provides a general de- scription of the future directions that can be pursued after this research. This chapter will evaluate the results of this research against its aim and objectives as well as revisit the main contributions of this research. It will also restate the main findings of this research as well as some major discussions that were made.