Classification of forensic autopsy reports through conceptual graph-based document representation model.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Forensic Pathology

Background:

• Text categorization is widely used for clinical reports, but traditional methods like bag-of-words (BoW) are insufficient for complex forensic autopsy reports.
• BoW fails to capture nuances like word inversion, synonymy, and polysemy, leading to low accuracy and robustness in classifying autopsy data.

Purpose of the Study:

• To develop an effective conceptual graph-based document representation (CGDR) technique for classifying forensic autopsy reports.
• To improve the accuracy and robustness of text classification for determining manner of death (MoD) and cause of death (CoD).

Main Methods:

• Developed a CGDR technique using term-based and SNOMED CT-based conceptual features extracted and represented as graphs.
• Trained a two-level classifier: the first level predicted MoD, and the second predicted CoD using CGDR.
• Compared CGDR against six state-of-the-art document representation techniques and evaluated one-level versus two-level classification.

Main Results:

• The CGDR technique achieved a 12% to 15% improvement in accuracy compared to baseline document representation methods.
• Two-level classification yielded better results than one-level classification for autopsy report analysis.
• The proposed CGDR technique demonstrated significant improvements in classifying forensic autopsy reports.

Conclusions:

• The CGDR technique offers a more accurate and robust approach to classifying forensic autopsy reports than traditional methods.
• The two-level classification strategy enhances the effectiveness of determining MoD and CoD.
• This system can serve as a valuable tool for pathologists, aiding in second opinions and improving the determination of cause of death.