An interpretable schizophrenia diagnosis framework using machine learning and explainable artificial intelligence

Figures & data

Table 1. Description of the features.

Sl. no	Feature name	Description	Sl. no	Feature name	Description	Sl. no	Feature name	Description
1	Participant	Control or Schizophrenia patient	18	Finger ridge count left index finger	No. of line present on the skin’s surface of the left index finger	35	Triradius right little finger	Three-ridge junction point on the right little finger
2	Sex	Gender of the participant	19	Triradius left index finger	Three-ridge junction point on the left index finger	36	Total finger ridge count left hand	Total no. of line present on the skin’s surfaces of all the finger in the left hand
3	Age range	Age range of the participants	20	Finger ridge count left middle finger	No. of line present on the skin’s surface of the left middle finger	37	Total finger ridge count right hand	Total no. of line present on the skin’s surfaces of all the finger in the right hand
4	Finger pattern left thumb	Finger pattern on the left thumb	21	Triradius left middle finger	Three-ridge junction point on the middle finger	38	Total finger ridge count	Total no. of lines present on the skin’s surfaces of all the fingers
5	Finger pattern left index finger	Finger pattern on the left index finger	22	Tinger ridge count left ring finger	No. of line present on the skin’s surface of the left ring finger	39	Total triradii left hand	Total three-ridge junction point in the left hand
6	Finger pattern left middle finger	Finger pattern on the left middle finger	23	Triradius left ring finger	Three-ridge junction point on the left ring finger	40	Total triradii right hand	Total three-ridge junction point in the right hand
7	Finger pattern left ring finger	Finger pattern on the left ring finger	24	Finger ridge count left little finger	No. of line present on the skin’s surface of the left little finger	41	Total triradii	Total three-ridge junction point on the palm or the finger
8	Finger pattern left little finger	Finger pattern on the left little finger	25	Triradius left little finger	Three-ridge junction point on the left little finger	42	a-b ridge count left hand	Count of ridges between two specific points a-b on the left hand
9	Finger pattern right thumb	Finger pattern on the right thumb	26	Finger ridge count right thumb	No. of line present on the skin’s surface of the right thumb	43	a-b ridge count right hand	Count of ridges between two specific points a-b on the right hand
10	Finger pattern right index finger	Finger pattern on the right index finger	27	Triradius right thumb	Three-ridge junction point on the right thumb	44	Total a-b ridge count	Total count of ridges between two specific points a-b
11	Finger pattern right middle finger	Finger pattern on the right middle finger	28	Finger ridge count right index finger	No. of line present on the skin’s surface of the right index finger	45	Palmar pattern left thenar region	Unique palm pattern near left thumb.
12	Finger pattern right ring finger	Finger pattern on the right ring finger	29	Triradius right index finger	Three-ridge junction point on the right index finger	46	Palmar pattern left hypothenar region	Unique palm pattern near base of left little finger
13	Finger pattern right little finger	Finger pattern on the right little finger	30	Finger ridge count right middle finger	No. of line present on the skin’s surface of the right middle finger	47	Palmar pattern right thenar region	Unique palm pattern near right thumb.
14	Palmar crease left hand	Palm lines on the left hand	31	Triradius right middle finger	Three-ridge junction point on the right middle finger	48	Palmar pattern right hypothenar region	Unique palm pattern near base of right little finger
15	Palmar crease right hand	Palm lines on the right hand	32	Finger ridge count right ring finger	No. of line present on the skin’s surface of the right ring finger	49	atd angle left hand/°	Angle measurement between the A (axial triradius), T (thenar triradius), and D (digital triradius) points on the palm of the left hand
16	Finger ridge count left thumb	No. of line present on the skin’s surface of the left thumb	33	Triradius right ring finger	Three-ridge junction point on the right ring finger	50	atd angle right hand/°	Angle measurement between the A (axial triradius), T (thenar triradius), and D (digital triradius) points on the palm of the right hand
17	Triradius of left thumb	Three-ridge junction point on the left thumb	34	Finger ridge count right little finger	No. of line present on the skin’s surface of the right little finger

Figure 1. Heatmap of the Pearsons correlation.

Figure 2. The features are ranked in relevance by the Mutual Information Algorithm.

Table 2. Methods and the features.

Methods	Important features
Pearsons’s Correlation	‘age range’, ‘sex_14’, sex_13’, ‘palmar pattern left thenar region_12’, ‘a-b ridge count left hand’, ‘finger pattern left thumb _8’, ‘atd angle right hand/°’, finger pattern right index finger_5’, ‘finger pattern right thumb_2’, ‘finger pattern right index finger_6’, palmar pattern left thenar region_11’, ‘finger pattern left thumb _2’, ‘palmar pattern left hypothenar region_10’, ‘total a-b ridge count’, ‘palmar crease left hand’, ‘finger pattern right middle finger_6’, ‘palmar pattern right hypothenar region_11’, ‘palmar pattern right hypothenar region_10’, ‘finger pattern left index finger_8’, ‘finger pattern left index finger_6’.
Mutual Information	'participant’, ‘age range’,'sex’,'finger pattern left index finger’,'total finger ridge count’,'atd angle right hand/°’,'triradius right ring finger’,'total triradii’,'palmar pattern left thenar region’,'triradius right index finger’,'triradius left little finger’,'finger pattern right index finger’,'finger pattern left little finger’,'finger ridge count right index finger’,'finger pattern right middle finger’,'palmar pattern left hypothenar region’,'finger pattern right ring finger’,'finger pattern right little finger’,'total triradii right hand’,'triradius of left thumb’,'finger ridge count left ring finger’.

Figure 3. The Visual representation of customized stacking

Figure 4. Represents the Flow chat of the voting methodology used in the study.

Figure 5. The pipeline of the machine learning methodology used in the study.

Table 3. All the classifiers with its accuracy, precision, recall and F1 score.

Models	Accuracy	Precision	Recall	F1 – score
Random forest	0.83	0.83	0.80	0.81
Logistic Regression	0.86	0.85	0.84	0.84
Decision Tree	0.83	0.83	0.80	0.81
KNN	0.66	0.62	0.59	0.58
SVM – linear kernel	0.86	0.85	0.84	0.84
SVM – sigmoid kernel	0.51	0.38	0.42	0.39
Stack 1	0.83	0.82	0.82	0.82
AdaBoost	0.83	0.82	0.82	0.82
CatBoost	0.83	0.82	0.82	0.82
LGBM	0.83	0.82	0.82	0.82
XGB	0.80	0.79	0.78	0.78
Stack 2	0.80	0.79	0.78	0.78
Final Stack	0.80	0.79	0.78	0.78
Hard voting	0.85	0.87	0.82	0.84
Soft voting	0.80	0.80	0.76	0.77
C4.5 (Decision Tree)	0.82	0.83	0.80	0.81
CART	0.82	0.82	0.82	0.82
Ridge	0.86	0.85	0.84	0.84

Figure 6. AUC curves for (a) Logistic regression, (b) SVM – linear kernel, (c) Ridge classifier

Figure 7. Precision-recall (PR) (a) Logistic regression, (b) SVM – Linear kernel, (c) Ridge classifier

Figure 8. Confusion matrix of (a) Logistic regression, (b) SVM – linear kernel, (c) Ridge Classifier

Figure 9. Beeswarm plot for SHAP (a) Beeswarm plot of 20 features (b) SHAP explainers for the 20 features

Figure 10. Attributes contributing for schizophrenic and not are represented by LIME.

Figure 11. Model predictions explained with QGraph.

Figure 12. Eli5 to explain model predictions.

Table 4. Five patients with schizophrenia and five patients without schizophrenia had anchor explanations.

Instance	Patient prediction	Anchor condition	Precision	Coverage
1	Control	age range < =  16.00 AND atd angle right hand/° > 42.62	1.00	0.34
2	Control	age range > 16.00 AND total triradii > 10.00	0.90	0.30
3	Control	age range < =  16.00 AND finger pattern right ring finger < =  7.00	1.00	0.21
4	Control	age range < =  16.00 AND finger ridge count left ring finger > 15.00	1.00	0.28
5	Control	age range < =  16.00 AND atd angle right hand/° > 42.62	0.99	0.35
6	Schizophrenic	age range > 16.00 AND triradius of left thumb < =  1.00	0.89	0.29
7	Schizophrenic	age range > 16.00 AND finger ridge count right index finger > 15.00	0.87	0.10
8	Schizophrenic	age range > 16.00 AND total triradii > 10.00	0.90	0.29
9	Schizophrenic	age range > 18.00 AND triradius of left thumb < =  1.00	0.96	0.11
10	Schizophrenic	age range > 18.00 AND atd angle right hand/° > 45.53	0.95	0.03

Table 5. Comparison between our proposed model and the existing models.

Sl.no	Model	Classifiers	Accuracy	Explainable AI Techniques
1	(Riyantoko & Diyasa, 2021)	SVM, XGBoost and AdaBoost	93%	SHAP
2	(Kawakura et al., 2022)	Trees, Logistic regression, Random forest, KNN, Gradient boosting, Naive Bayes and SVM	77.6%	No
3	(Gallagher et al., 2017)	Random forest, Logistic regression, MPL, AdaBoost, SVM radial, XGBoost	81.7%	No
4	Our proposed model	Random forests, Decision trees, Logistic regression, KNN, SVM (Linear, sigmoidal), AdaBoost, LGBM, CatBoost, XGBoost, Hard and soft voting, C4.5, CART, Ridge and stacking models	86%	SHAP, LIME, QLattice, Anchor, ELI5

Riyantoko, P. A., & Diyasa, I. G. S. M. (2021). October). “FQAM” Feyn-QLattice Automation Modelling: Python Module of Machine Learning for Data Classification in Water Potability. In 2021 International Conference on Informatics, Multimedia, Cyber and Information System (ICIMCIS (pp. 135-141). IEEE.

 Google Scholar

Kawakura, S., Hirafuji, M., Ninomiya, S., & Shibasaki, R. (2022). Adaptations of explainable artificial intelligence (XAI) to agricultural data models with ELI5. PDPbox, and skater using diverse agricultural worker data. European Journal of Artificial Intelligence and Machine Learning, 1(3), 27–34.

 Google Scholar

Gallagher, R. J., Reing, K., Kale, D., & Ver Steeg, G. (2017). Anchored correlation explanation: Topic modeling with minimal domain knowledge. Transactions of the Association for Computational Linguistics, 5, 529–542.

 Google Scholar

Data availability statement

Data will be made available after obtaining prior permission from the corresponding author.