1. Demographic information and Baseline characteristics of the study subjects
When analyzing patients based on outcomes, specifically mortality, the study included 240 patients, of which 163 (67.9%) were male and 77 (32.1%) were female, with 38 (15.8%) fatalities. Table 1 contrasts the baseline characteristics of non-death and death groups, revealing that the median age was significantly higher in the death group (p < 0.001), and that patients with severe COVID-19 were more likely to succumb (p < 0.001). The fasting blood glucose (Glu) level was also higher in the death group compared to the non-death group (p < 0.05). Statistical differences were observed in levels of C-reactive protein (CRP), glutamic oxaloacetic transaminase (AST), creatinine (CREA), high-sensitivity troponin T (hs-cTnT), creatine kinase (CK), and lactate dehydrogenase (LDH) (p < 0.05). Table 2 compares non-severe and severe groups, showing that patients in the severe group were older (p < 0.001). Levels of CRP, erythrocyte sedimentation rate (ESR), PCT, CREA, LDH, and Glu were higher in the severe group compared to the non-severe group (p < 0.05). Levels of ALB was higher in the non-severe group compared to the severe group (p < 0.05).
Table 1. Baseline demographic and clinical characteristics of patients included in the non-death and death groups.
Baseline characteristics
|
Non-death group
|
Death group
|
p value
|
Number of people
|
202
|
38
|
|
Age (median [IQR])
|
66.00 [55.00, 74.75]
|
80.00 [71.00, 84.25]
|
< 0.001
|
CRP (median [IQR])
|
75.80 [23.60, 118.50]
|
124.00 [57.98, 171.50]
|
0.011
|
Severe case (%)
|
142 (70.3)
|
10 (26.3)
|
< 0.001
|
|
60 (29.7)
|
28 (73.7)
|
|
ALT (median [IQR])
|
23.60 [14.80, 39.40]
|
20.65 [15.82, 41.40]
|
0.938
|
AST (median [IQR])
|
26.70 [18.75, 45.65]
|
35.80 [25.38, 64.00]
|
0.012
|
ESR (median [IQR])
|
50.00 [28.75, 74.00]
|
60.00 [34.00, 80.00]
|
0.274
|
PCT (median [IQR])
|
0.31 [0.10, 1.62]
|
0.62 [0.16, 2.44]
|
0.146
|
CREA
(median [IQR])
|
87.00 [69.00, 164.00]
|
139.55 [77.25, 268.00]
|
0.042
|
LDH (median [IQR])
|
287.60 [229.50, 415.72]
|
368.50 [302.00, 497.00]
|
0.002
|
CK (median [IQR])
|
92.50 [44.75, 193.50]
|
180.00 [59.10, 366.48]
|
0.026
|
hs-cTnT
(median [IQR])
|
21.55 [11.97, 94.75]
|
46.10 [21.96, 92.30]
|
0.030
|
β-HB
(median [IQR])
|
0.66 [0.38, 1.60]
|
1.02 [0.68, 1.40]
|
0.080
|
HbA1c
(median [IQR])
|
8.42 [7.18, 10.60]
|
7.80 [7.30, 9.38]
|
0.401
|
Glu (median [IQR])
|
9.50 [7.00, 13.00]
|
12.50 [9.85, 17.80]
|
0.003
|
ALB (mean (SD))
|
32.66 (5.91)
|
32.63 (6.97)
|
0.978
|
TP (mean (SD))
|
60.59 (7.69)
|
62.26 (9.57)
|
0.241
|
Table 2. Baseline demographic and clinical characteristics of patients included in the non-severe and severe groups.
Baseline characteristics
|
Non-severe group
|
Severe group
|
p value
|
Number of people
|
154
|
88
|
|
Age (median [IQR])
|
65.00 [52.25, 76.75]
|
72.50 [63.75, 79.25]
|
< 0.001
|
CRP (median [IQR])
|
56.55 [16.38, 100.61]
|
103.00 [68.75, 151.75]
|
< 0.001
|
ALT (median [IQR])
|
23.10 [14.40, 37.90]
|
24.00 [15.48, 42.72]
|
0.561
|
AST (median [IQR])
|
27.10 [19.40, 47.40]
|
28.95 [20.70, 48.33]
|
0.607
|
ESR (median [IQR])
|
43.00 [23.00, 69.00]
|
62.00 [36.75, 93.75]
|
0.001
|
PCT (median [IQR])
|
0.25 [0.08, 1.50]
|
0.50 [0.15, 2.74]
|
0.016
|
CREA
(median [IQR])
|
84.00 [67.00, 144.00]
|
131.00 [78.00, 277.00]
|
< 0.001
|
LDH (median [IQR])
|
270.00 [219.25, 415.90]
|
338.50 [280.35, 475.02]
|
0.001
|
CK (median [IQR])
|
85.00 [45.00, 194.98]
|
125.00 [51.00, 298.00]
|
0.06
|
hs-cTnT
(median [IQR])
|
21.85 [12.10, 102.25]
|
28.75 [14.88, 80.38]
|
0.355
|
β-HB
(median [IQR])
|
0.70 [0.42, 1.77]
|
0.78 [0.40, 1.34]
|
0.822
|
HbA1c
(median [IQR])
|
8.30 [7.00, 10.59]
|
8.46 [7.32, 10.47]
|
0.272
|
Glu (median [IQR])
|
9.15 [6.82, 12.75]
|
11.75 [8.07, 16.47]
|
0.005
|
ALB (mean (SD))
|
33.56 (6.08)
|
31.17 (5.74)
|
0.003
|
TP (mean (SD))
|
61.54 (7.88)
|
59.65 (8.12)
|
0.082
|
p value < 0.05 indicates statistical significance.
Abbreviations: SD, Standard deviation; IQR, Interquartile Range.
2. Model feature filtering and The best model
We compared the internal 5-fold cross-validation results of five machine learning models across three groups, evaluating the AUC, sensitivity, specificity, and accuracy (Tables 3 and 4). The LR model demonstrated the highest accuracy (0.816) and AUC (0.933) in predicting mortality among COVID-19 patients with DKA (Table 3, Figure 1a). Additionally, the LR model achieved the highest accuracy (0.875) and AUC (0.898) in predicting the progression of patients to severe illness (Table 4, Figure 1b). Thus, the LR model was selected as the most effective algorithm for predicting patient mortality.
In our study, the LR algorithm was employed to evaluate all examination items, continuously refining the combination of items and diagnostic performance. Eventually, age, COVID-19 classification, urinary occult blood (BLD), urinary ketone bodies (U-Ket), AST, TP, and body temperature (BT) were identified as key factors for constructing the optimal machine learning model to predict mortality rates in COVID-19 patients with DKA. Furthermore, the LR algorithm identified five crucial examination items—COVID-19 classification, BLD, hs-cTnT, myoglobin (Mb), and PCT—for predicting the likelihood of patient condition worsening.
Table 3. Summary of prediction results of multiple models after 5-fold cross-validation for the non-death and death groups.
Model
|
AUC
|
Sensitivity
|
Specificity
|
Precision
|
Accuracy
|
LR
|
0.933
|
1.000
|
0.780
|
0.471
|
0.816
|
XGB
|
0.873
|
0.875
|
0.780
|
0.438
|
0.796
|
RF
|
0.902
|
1.000
|
0.707
|
0.400
|
0.755
|
SVM
|
0.864
|
0.875
|
0.800
|
0.467
|
0.813
|
MLP
|
0.746
|
0.857
|
0.525
|
0.240
|
0.575
|
Table 4. Summary of prediction results of multiple models after 5-fold cross-validation for the non-severe and severe groups.
Model
|
AUC
|
Sensitivity
|
Specificity
|
Precision
|
Accuracy
|
LR
|
0.898
|
0.895
|
0.793
|
0.739
|
0.875
|
XGB
|
0.896
|
0.895
|
0.862
|
0.810
|
0.833
|
RF
|
0.880
|
0.790
|
0.862
|
0.790
|
0.833
|
SVM
|
0.773
|
1.000
|
0.448
|
0.543
|
0.667
|
MLP
|
0.715
|
0.700
|
0.667
|
0.583
|
0.680
|
3. Interpretation and Evaluation of machine learning models
Based on the SHAP algorithm, using the LR model to predict the mortality of COVID-19 patients with DKA shows (Figure 2a) that age, COVID-19 classification, urinary occult blood, U-Ket, AST, TP, and BT have a significant impact on the prediction results, with age having the greatest impact on predicting outcomes. Generally, age, COVID-19 classification, BLD, and AST are positively correlated with mortality and are considered risk factors. When predicting the progression of COVID-19 patients with DKA to severe illness (Figure 2b) using the LR model, it was found that COVID-19 classification, BLD, hs-cTnT, hematocrit, Mb, percentage of monocytes (MONO%), and PCT have a significant prognostic impact. Among these, COVID-19 classification has the most substantial effect on the outcomes. Overall, factors such as COVID-19 classification, BLD, hs-cTnT, Mb, and PCT are positively associated with progression to severe illness, marking them as risk factors.