Predicting chemotherapy-induced myelosuppression in colorectal cancer: An interpretable, machine learning-based nomogram

Table 1 Baseline characteristics of colorectal cancer patients, n (%)

Characteristic	Overall (n = 765)	Training set (n = 612)	Validation set (n = 153)	P value
Age, median (95%CI)	62.26 (61.58-62.96)	62.12 (61.36-62.88)	62.84 (61.16-64.51)	0.422
BSA, median (95%CI)	1.71 (1.69-1.72)	1.71 (1.69-1.72)	1.71 (1.68-1.74)	0.996
BMI, median (95%CI)	22.94 (22.70-23.17)	22.99 (22.72-23.26)	22.75 (22.23-23.27)	0.428
ALB, median (95%CI)	37.53 (37.27-37.80)	37.48 (37.19-37.50)	37.74 (37.10-38.37)	0.453
CEA, median (95%CI)	182.30 (125.72-246.75)	178.10 (115.79-240.41)	199.10 (53.75-344.45)	0.770
CA19-9, median (95%CI)	169.03 (113.66-237.21)	177.95 (108.12-247.78)	133.33 (0.11-266.55)	0.571
CA125, median (95%CI)	22.20 (19.20-25.77)	21.93 (18.30-25.56)	23.30 (13.48-33.12)	0.804
Gender				0.607
Male	449 (58.7)	362 (59.15)	87 (56.86)
Female	316 (41.3)	250 (40.85)	66 (43.14)
Smoking				0.572
Yes	275 (35.9)	223 (36.44)	52 (33.99)
No	316 (41.3)	389 (63.56)	101 (66.01)
Diabetes				0.529
Yes	128 (16.7)	105 (17.16)	23 (15.03)
No	637 (83.2)	507 (82.84)	130 (84.97)
Hypertension				0.417
Yes	307 (40.1)	250 (40.85)	57 (37.25)
No	458 (59.8)	362 (59.15)	96 (62.75)
T				0.491
1	18 (2.4)	15 (2.45)	3 (1.96)
2	84 (11.0)	67 (10.95)	17 (11.11)
3	264 (34.5)	323 (52.78)	76 (49.67)
4	399 (52.2)	207 (33.82)	57 (37.25)
N				0.255
0	4 (0.5)	126 (20.59)	36 (23.53)
1	162 (21.2)	261 (42.65)	68 (44.44)
2	270 (35.3)	222 (36.27)	48 (31.37)
3	329 (43.0)	3 (0.49)	1 (0.65)
M				0.103
0	355 (46.4)	293 (47.88)	62 (40.52)
1	410 (53.5)	319 (52.12)	91 (59.48)
Staging				0.180
I	15 (2.0)	11 (1.8)	4 (2.61)
II	66 (8.6)	56 (9.15)	10 (6.54)
III	267 (34.9)	219 (35.78)	48 (31.37)
IV	417 (54.5)	326 (53.27)	91 (59.48)
Position				0.687
Ascending colon	21 (2.7)	118 (19.28)	30 (19.61)
Transverse colon	27 (3.5)	15 (2.45)	6 (3.92)
Descending colon	148 (19.3)	24 (3.92)	3 (1.96)
Sigmoid colon	209 (27.3)	164 (26.8)	45 (29.41)
Rectum	360 (47.1)	291 (47.55)	69 (45.1)
Hepatic metastasis				0.195
Yes	261 (34.1)	202 (33.01)	59 (38.56)
No	504 (65.8)	410 (66.99)	94 (61.44)
Lung metastasis				0.064
Yes	214 (27.9)	162 (26.47)	52 (33.99)
No	551 (72.0)	450 (73.53)	101 (66.01)
Peritoneum metastasis				0.201
Yes	47 (6.1)	41 (6.7)	6 (3.92)
No	718 (93.8)	571 (93.3)	147 (96.08)
KRAS				0.961
Yes	124 (16.2)	99 (16.18)	25 (16.34)
No	641 (83.7)	513 (83.82)	128 (83.66)
BRAF				0.051
Yes	15 (1.9)	15 (2.45)	0 (0)
No	750 (98)	597 (97.55)	153 (100)
TP53				0.286
Yes	61 (7.9)	52 (8.5)	9 (5.88)
No	704 (92)	560 (91.5)	144 (94.12)
Myelosuppression				0.336
Yes	250 (32.6)	195 (31.86)	55 (35.95)
No	515 (67.3)	417 (68.14)	98 (64.05)
Chemotherapy cycles				0.572
1-2 cycles	315 (41.2)	258 (42.16)	57 (37.25)
3-4 cycles	237 (31.0)	182 (29.74)	55 (35.95)
5 or more cycles	213 (27.8)	172 (28.1)	41 (26.8)
Chemotherapy regimens				0.248
CapeOx	571 (74.6)	463 (75.65)	108 (70.59)
FOLFOX	85 (11.1)	63 (10.29)	22 (14.38)
FOLFIRI	109 (14.2)	86 (14.05)	23 (15.03)

BSA: Body surface area; BMI: Body mass index; ALB: Albumin; CEA: Carcinoembryonic antigen; CA: Carbohydrate antigen; CI: Confidence interval; T: Tumor; N: Node; M: Metastasis.

Table 2 Performance of 10 machine learnings for predicting myelosuppression after first-line chemotherapy for colorectal cancer

MLs	AUC (95%CI)	AUPRC (95%CI)	Accuracy (%)	Sensitivity (%)	Specificity (%)	F1 score	PPV	NPV
Training set
Adaptive boosting	0.88 (0.85-0.91)	0.79 (0.73-0.84)	0.79	0.49	0.94	0.61	0.81	0.78
Artificial neural network	0.96 (0.95-0.97)	0.94 (0.91-0.96)	0.73	0.53	0.83	0.57	0.61	0.78
Decision tree	1.00 (1.00-1.00)	1.00 (1.00-1.00)	1.00	1.00	1.00	1.00	1.00	1.00
Extra trees	1.00 (1.00-1.00)	1.00 (1.00-1.00)	1.00	1.00	1.00	1.00	1.00	1.00
Gradient boosting machine	0.99 (0.99-1.00)	0.99 (0.99-1.00)	0.98	0.93	0.99	0.96	0.99	0.97
K-Nearest neighbors	0.91 (0.89-0.93)	0.79 (0.75-0.84)	0.90	0.85	0.92	0.85	0.85	0.92
Logistic regression	0.75 (0.71-0.79)	0.57 (0.50-0.65)	0.69	0.25	0.92	0.35	0.61	0.70
Random forest	1.00 (0.99-1.00)	1.00 (0.99-1.00)	1.00	1.00	1.00	1.00	1.00	1.00
Support vector machine	0.87 (0.83-0.90)	0.79 (0.73-0.84)	0.68	0.05	1.00	0.10	1.00	0.67
Extreme gradient boosting	1.00 (0.99-1.00)	1.00 (0.99-1.00)	1.00	1.00	1.00	1.00	1.00	1.00
Validation set
Adaptive boosting	0.83 (0.76-0.89)	0.72 (0.60-0.85)	0.78	0.35	0.95	0.47	0.71	0.79
Artificial neural network	0.69 (0.60-0.78)	0.61 (0.48-0.74)	0.65	0.42	0.75	0.40	0.39	0.77
Decision tree	0.70 (0.62-0.78)	0.52 (0.41-0.64)	0.82	0.81	0.83	0.72	0.65	0.92
Extra trees	0.94 (0.89-0.97)	0.90 (0.82-0.96)	0.87	0.72	0.93	0.76	0.79	0.89
Gradient boosting machine	0.92 (0.86-0.97)	0.90 (0.84-0.95)	0.83	0.67	0.89	0.69	0.71	0.88
K-Nearest neighbors	0.75 (0.67-0.83)	0.62 (0.48-0.74)	0.80	0.70	0.85	0.67	0.64	0.88
Logistic regression	0.67 (0.59-0.76)	0.49 (0.38-0.65)	0.68	0.20	0.86	0.27	0.38	0.74
Random forest	0.96 (0.93-0.98)	0.93 (0.87-0.97)	0.88	0.72	0.95	0.78	0.84	0.90
Support vector machine	0.75 (0.67-0.83)	0.65 (0.52-0.77)	0.75	0.12	1.00	0.21	1.00	0.74
Extreme gradient boosting	0.97 (0.94-0.99)	0.92 (0.83-0.99)	0.88	0.79	0.92	0.79	0.79	0.92

AUC: Area under the curve; ML: Machine learning; AUPRC: Area under the precision-recall curve; CI: Confidence interval; PPV: Positive predictive value; NPV: Negative predictive value.

Table 3 Logistic regression based on 10 machine learnings for predicting myelosuppression

MLs	Univariate logistic regression		Multivariate logistic regression
	OR (95%CI)	P value	OR (95%CI)	P value
Adaptive boosting	49.558 (15.807-155.376)	0.000	0.083 (0.003-2.486)	0.151
Artificial neural network	54.444 (13.032-227.465)	0.000	1.572 (0.043-57.273)	0.805
Decision tree	10.587 (5.084-22.047)	0.000	0.563 (0.102-3.109)	0.510
Extra trees	390.471 (75.734-2013.192)	0.000	31.948 (2.468-413.586)	0.008
Gradient boosting machine	104.831 (32.579-337.322)	0.000	2.169 (0.176-26.731)	0.546
K-Nearest neighbors	24.992 (9.652-64.711)	0.000	3.081 (0.419-22.650)	0.269
Logistic regression	279.116 (24.997-3116.614)	0.000	0.225 (0.000-129.448)	0.646
Random forest	404.139 (88.973-1835.710)	0.000	94.621 (1.178-7597.788)	0.042
Support vector machine	4.180 (0.332-52.698)	0.269	23.780 (0.434-1303.403)	0.121
Extreme gradient boosting	131.875 (39.641-438.709)	0.000	2.669 (0.170-41.930)	0.485

ML: Machine learning; CI: Confidence interval; OR: Odds ratio.

Table 4 Performance comparison of extreme gradient boosting, clinic nomogram, and clinic-machine learning nomograms

Model	AUC (95%CI)	AUPRC (95%CI)	Accuracy (%)	Sensitivity (%)	Specificity (%)	F1 score	PPV	NPV
Training set
Extreme gradient boosting	1.00 (1.00-1.00)	1.00 (0.99-1.00)	1.00	1.00	1.00	1.00	1.00	1.00
Clinic nomogram	0.68 (0.63-0.72)	0.49 (0.43-0.56)	0.69	0.25	0.92	0.69	0.61	0.71
Clinic-machine learning nomogram	0.99 (0.99-1.00)	0.99 (0.99-1.00)	1.00	1.00	1.00	1.00	1.00	1.00
Validation set
Extreme gradient boosting	0.97 (0.94-0.99)	0.92 (0.83-0.99)	0.88	0.79	0.92	0.79	0.79	0.92
Clinic nomogram	0.61 (0.52-0.71)	0.48 (0.36-0.62)	0.68	0.21	0.86	0.27	0.38	0.74
Clinic-machine learning nomogram	0.96 (0.92-0.98)	0.93 (0.87-0.98)	0.90	0.77	0.95	0.80	0.85	0.91

AUC: Area under the curve; AUPRC: Area under the precision-recall curve; CI: Confidence interval; PPV: Positive predictive value; NPV: Negative predictive value.

Table 5 Performance of clinic-machine learning nomogram in the testing set

Model	AUC (95%CI)	AUPRC (95%CI)	Accuracy (%)	Sensitivity (%)	Specificity (%)	F1 score	PPV	NPV
Clinic-machine learning nomogram	0.95 (0.93-0.99)	0.83 (0.71-0.92)	0.81	0.52	0.97	0.65	0.89	0.79

AUC: Area under the curve; AUPRC: Area under the precision-recall curve; CI: Confidence interval; PPV: Positive predictive value; NPV: Negative predictive value.