Objective

2

To construct a tongue manifestation-based diabetic foot risk prediction model using LASSO regression and Bayesian network analysis.

Methods

2

The clinical data of patients with type 2 diabetes （T2D） and diabetic foot admitted to the Second Teaching Hospital of Tianjin University of Traditional Chinese Medicine from January 2019 to October 2020 were collected， and 27 variables including age， gender， body mass index （BMI）， smoking， education level， light red tongue， pale tongue， red crimson tongue， light purple tongue， dusky red tongue， red tongue tip， enlarged tongue， red tongue edge， thin tongue， spotted tongue， teeth-marked tongue， cracked tongue， thin coating， white coating， yellow coating， scanty coating， peeled coating， glossy coating， thick coating， greasy coating， fasting blood glucose （FBG） and 2-hour postprandial blood glucose （2hPBG）were analyzed. Logistic univariate analysis was used to calculate the unadjusted OR values （cOR values） of tongue manifestation-related variables. The influencing factors significantly related to the occurrence of diabetic foot were screened out by LASSO regression， and then the interaction between variables was further analyzed and reasoned by Bayesian network analysis. Finally， the predictive effectiveness of the model was evaluated.

Results

2

A total of 293 diabetic foot patients （diabetic foot group） and 4430 T2D patients （T2D group） were included. There were significant differences in age， BMI， gender， smoking， and education level between the two groups （

P

＜0.05）， with no significant differences in FBG and 2hPBG levels （

P

＞0.05）. The risk of diabetic foot was increased in patients with light purple tongue （cOR = 1.956） and thin tongue （cOR = 4.315） through Logistic univariate analysis. Six variables that were most significantly associated with the occurrence of diabetic foot were screened out through LASSO regression， including thin tongue， smoking， dusky red tongue， greasy coating， thick coating， and white coating. Bayesian network analysis further found that thin tongue， dusky red tongue， greasy coating， thick coating， and white coating were directly associated with the occurrence of diabetic foot， while smoking was indirectly linked. The prediction model showed that the max probability of developing diabetic foot was 7.77% in patients with thick coating， and that was 6.43% when there was greasy coating. The probability of thick coating and greasy coating were higher in patients who smoked than those who not， while the probability of dusky red tongue was higher in non-smoking patients. The model efficacy evaluation showed accuracy of 0.93 and 0.97 of f1 value， indicating a good prediction effectiveness.

Conclusion

2

The tongue manifestation-based diabetic foot risk prediction model constructed with LASSO regression and Bayesian network ana-lysis shows that the probability of diabetic foot in T2D patients with thick coating and greasy coating is 7.77% and 6.43%， respectively， and the probability may further increase if the patients smoke.