A case-base fuzzification process: diabetes diagnosis case study

Abstract

Medical case-based reasoning (CBR) systems require the handling of vague or imprecise data. The fuzzy set theory is particularly suitable for this purpose. This paper proposes a case-base preparation framework for CBR systems, which converts the electronic health record medical data into fuzzy CBR knowledge. It generates fuzzy case-base knowledge by suggesting a standard crisp entity–relationship data model for CBR case-base. The resulting data model is fuzzified using a proposed relational data model fuzzification methodology. The performances of this methodology and its resulting fuzzy case-base structure are evaluated. Diabetes diagnosis is used as a case study. A set of 60 real diabetic cases is used in the study. A fuzzy CBR system is implemented to check the diagnoses accuracy. It combines the resulting fuzzy case-base with a proposed fuzzy similarity measure. Experimental results indicate that the proposed fuzzy CBR method is superior to traditional CBR and other machine-learning methods. Our fuzzy CBR achieves an accuracy of 95%, a precision of 96%, a recall 97.96%, an f-measure of 96.97%, a specificity of 81.82%, and good robustness for dealing with vagueness. The resulting fuzzy case-base relational database enhances the representation of case-base knowledge, the performance of retrieval algorithms, and the querying capabilities of CBR systems.

Appendices

Appendix A

The performance evaluation of the proposed fuzzy CBR system

$K=3$
Query cases	Retrieved cases	FKI-CBR diagnose								Domain expert diagnose
		Diabetes diagnosis	Nephropathy	Hypercholesterolemia	Cancer	Liver	Glomerulonephritis	Splenomegaly	Confidence (%)
Case 1	Case 1-1	D	N	N	N	N	N	N	90	Diabetes Diagnosis: D
	Case 1-2	P	N	N	N	N	N	N	88.9	Percentage of similarity for another diagnosis: 9/10
	Case 1-3	D	N	H	N	N	N	N	88.8
Case 2	Case 2-1	N	N	N	N	N	N	N	95	Diabetes Diagnosis: N
	Case 2-2	N	N	N	N	N	N	N	91	Percentage of similarity for another diagnosis: 10/10
	Case 2-3	N	N	N	N	N	N	N	90
Case 3	Case 3-1	P	SH	N	N	N	G	N	88	Diabetes Diagnosis: D
	Case 3-2	DG	SH	N	N	F	G	N	85	Percentage of similarity for another diagnosis: 10/10
	Case 3-3	D	N	N	N	N	N	N	84
Case 4	Case 4-1	D	N	H	N	N	G	N	94	Diabetes Diagnosis: D
	Case 4-2	D	N	H	N	N	G	N	93.2	Percentage of similarity for another diagnosis: 10/10
	Case 4-3	N	NP	N	N	N	G	N	86.5
Case 5	Case 5-1	D	N	N	N	N	N	N	100	Diabetes Diagnosis: D
	Case 5-2	D	N	N	O	N	N	N	88.8	Percentage of similarity for another diagnosis: 10/10
	Case 5-3	D	N	N	N	N	N	N	86.9
Case 6	Case 6-1	P	SH	N	N	N	G	N	81	Diabetes Diagnosis: P
	Case 6-2	D	N	N	N	N	N	N	79	Percentage of similarity for another diagnosis: 8/10
	Case 6-3	D	N	N	HCC	HCC	N	N	79
Case 7	Case 7-1	N	N	N	N	N	N	N	92.4	Diabetes Diagnosis: N
	Case 7-2	D	N	N	N	N	N	N	89.6	Percentage of similarity for another diagnosis: 10/10
	Case 7-3	N	N	N	N	N	N	N	89.4
Case 8	Case 8-1	D	N	N	N	HCC	N	N	99	Diabetes Diagnosis: D
	Case 8-2	N	N	H	N	N	N	N	82.5	Percentage of similarity for another diagnosis: 9/10
	Case 8-3	P	N	N	N	N	N	N	80
Case 9	Case 9-1	D	N	H	N	N	N	N	93	Diabetes Diagnosis: D
	Case 9-2	DG	SH	N	N	N	G	N	91	Percentage of similarity for another diagnosis: 10/10
	Case 9-3	D	N	N	N	HCC	N	SG	90
Case 10	Case 10-1	D	N	N	N	N	N	N	91	Diabetes Diagnosis: D
	Case 10-2	DG	N	N	N	N	N	N	89.9	Percentage of similarity for another diagnosis: 10/10
	Case 10-3	P	N	H	N	N	N	N	89.2

N normal, A abnormal, D diabetic, P pre-diabetic, H hypercholesterolemia, O ovarian cancer, DG gestational diabetes, SH Shrunken Kidney, G glomerulonephritis, NP nephropathy, SG splenomegaly, HCC hepatocellular carcinoma, F fatty liver, LC liver cirrhosis

Appendix B

A comparison of the proposed fuzzy CBR system and traditional CBR system

Query case	Domain expert decision	Proposed system decision	Traditional system decision
Case1	Diabetes diagnosis: D	D (90%)	P (78%)
Case 2	Diabetes diagnosis: N	N (95%)	N (83%)
Case 3	Diabetes diagnosis: D	P (88%)	D (70%)
Case 4	Diabetes diagnosis: D	D (94%)	P (64%)
Case 5	Diabetes diagnosis: D	D (100%)	D (80%)
Case 6	Diabetes diagnosis: P	P (81%)	D (64%)
Case 7	Diabetes diagnosis: N	N (92.4%)	N (85%)
Case 8	Diabetes diagnosis: D	D (99%)	P (79%)
Case 9	Diabetes diagnosis: D	D (93%)	P (74%)
Case 10	Diabetes diagnosis: D	D (91%)	P (75%)

Appendix C

A comparison of our FCBR and ML classifiers using our case-base data

Fold	Algorithm	Precision (%)	TPR-Recall (%)	Accuracy (%)	F-Measure (%)
Machine-learning algorithms
Twofold	C4.5	90	93.1	90	91.5
	k-NN ($k=3$)	80	69	66.66	74.1
	SVM	75.8	86.2	68.33	80.6
	Naive Bayes	86.2	86.2	75	86.2
	ANN	70.6	82.8	65	76.2
Threefold	C4.5	89.7	89.7	88.33	89.7
	k-NN ($k=3$)	74.1	69	60	71.4
	SVM	78.8	89.7	71.66	83.9
	Naive Bayes	83.9	89.7	75	86.7
	ANN	73.5	86.2	65	79.4
Fourfold	C4.5	89.7	89.7	88.33	89.7
	k-NN ($k=3$)	76.9	69	65	72.7
	SVM	77.4	82.8	71.66	80
	Naive Bayes	79.3	79.3	71.66	79.3
	ANN	76.7	79.3	66.66	78
Fivefold	C4.5	93.1	93.1	91.67	93.1
	k-NN ($k=3$)	73.1	65.5	63.33	69.1
	SVM	75	82.8	66.66	78.7
	Naive Bayes	83.3	69	66.66	75.5
	ANN	75.8	86.2	68.33	80.6
Sixfold	C4.5	89.7	89.7	88.33	89.7
	k-NN ($k=3$)	74.1	69	63.33	71.4
	SVM	81.3	89.7	75	85.2
	Naive Bayes	86.2	86.2	73.33	86.2
	ANN	76.5	89.7	70	82.5
Sevenfold	C4.5	90	93.1	90	91.5
	k-NN ($k=3$)	63.3	65.5	58.33	64.4
	SVM	76.7	79.3	71.66	78
	Naive Bayes	75.9	75.9	70	75.9
	ANN	76.7	79.3	65	78
Eightfold	C4.5	89.3	86.2	86.66	87.7
	k-NN ($k=3$)	78.6	75.9	66.66	77.2
	SVM	82.1	79.3	71.66	80.7
	Naive Bayes	78.1	86.2	73.33	82
	ANN	77.4	82.8	71.66	80
Ninefold	C4.5	89.7	89.7	88.33	89.7
	k-NN ($k=3$)	83.3	69	66.66	75.5
	SVM	80.6	86.2	76.66	83.3
	Naive Bayes	82.1	79.3	73.33	80.7
	ANN	83.9	89.7	75	86.7
Tenfold	C4.5	89.7	89.7	88.33	89.7
	k-NN ($k=3$)	73.1	65.5	56.67	69.1
	SVM	80.6	86.2	75	83.3
	Naive Bayes	83.3	86.2	75	84.7
	ANN	76.5	89.7	70	82.5
k-fold ($k=60$) $\equiv$ LOOCV	C4.5	89.7	89.7	88.33	89.7
	k-NN ($k=3$)	73.1	65.5	58.33	69.1
	SVM	80	82.8	75	81.4
	Naive Bayes	85.2	79.3	73.33	82.1
	ANN	78.8	89.7	75	83.9
Average (%)		80.568	82.086	73.0972	81.164
Maximum (%)		93.1	93.1	91.67	93.1
Conventional CBR system		76	47.5	55	58.46
Proposed FCBR system		96	97.96	95	96.97