Key messages
Background
Methods
Systematic review design, definitions, and inclusion/exclusion criteria
Search strategy
Eligibility criteria
Study selection
Review process and data extraction
Results
Identification of eligible studies
Purpose of machine learning in the ICU
Number of patients analysed | |||||||
---|---|---|---|---|---|---|---|
Aim of study | Number (%) of papers with this aima | < 100 | 100–1000 | 1000–10,000 | 10,000–100,000 | 100,000–1,000,000 | Number not reported |
Predicting complications | 79 (30.6%) | 23 (29.1%) | 26 (32.9%) | 17 (21.5%) | 8 (10.1%) | 3 (3.8%) | 2 (2.5%) |
Predicting mortality | 70 (27.1%) | 11 (15.7%) | 19 (27.1%) | 19 (27.1%) | 18 (25.7%) | 1 (1.4%) | 2 (2.9%) |
Improving prognostic models/risk scoring system | 43 (16.7%) | 8 (18.6%) | 16 (37.2%) | 8 (18.6%) | 8 (18.6%) | 2 (4.7%) | 1 (2.3%) |
Classifying sub-populations | 29 (11.2%) | 11 (37.9%) | 8 (27.6%) | 6 (20.7%) | 2 (6.9%) | 0 (0.0%) | 2 (6.9%) |
Alarm reduction | 21 (8.14%) | 9 (42.9%) | 5 (23.8%) | 7 (33.3%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) |
Predicting length of stay | 18 (6.98%) | 3 (16.7%) | 7 (38.9%) | 5 (27.8%) | 3 (16.7%) | 0 (0.0%) | 0 (0.0%) |
Predicting health improvement | 17 (6.59%) | 5 (29.4%) | 10 (58.8%) | 2 (11.8%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) |
Determining physiological thresholds | 16 (6.20%) | 10 (62.5%) | 4 (25.0%) | 1 (6.2%) | 0 (0.0%) | 0 (0.0%) | 1 (6.2%) |
Improving upon previous methods | 5 (1.94%) | 2 (40.0%) | 1 (20.0%) | 1 (20.0%) | 1 (20.0%) | 0 (0.0%) | 0 (0.0%) |
Detecting spurious recorded values | 3 (1.16%) | 1 (33.3%) | 2 (66.7%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) |
Total (accounting for duplicates) | 258 | 72 (27.9%) | 84 (32.6%) | 55 (21.3%) | 35 (13.6%) | 6 (2.33%) | 6 (2.33%) |
Type of machine learning
Number of patients analysed | |||||||
---|---|---|---|---|---|---|---|
Type of machine learning | Number (%) of papers with this typea | < 100 | 100–1000 | 1000–10,000 | 10,000–100,000 | 100,000–1,000,000 | Number not reported |
Neural network | 72 (42.6%) | 14 (19.4%) | 27 (37.5%) | 20 (27.8%) | 9 (12.5%) | 2 (2.8%) | 0 (0.0%) |
Support vector machine | 40 (23.7%) | 12 (30.0%) | 15 (37.5%) | 8 (20.0%) | 4 (10.0%) | 1 (2.5%) | 0 (0.0%) |
Classification/decision trees | 35 (20.7%) | 6 (17.1%) | 11 (31.4%) | 10 (28.6%) | 5 (14.3%) | 1 (2.9%) | 2 (5.7%) |
Random forest | 21 (12.4%) | 1 (4.8%) | 9 (42.9%) | 5 (23.8%) | 4 (19.0%) | 2 (9.5%) | 0 (0.0%) |
Naive Bayes/Bayesian networks | 19 (11.2%) | 4 (21.1%) | 5 (26.3%) | 6 (31.6%) | 2 (10.5%) | 1 (5.3%) | 1 (5.3%) |
Fuzzy logic/rough set | 12 (7.1%) | 3 (25.0%) | 5 (41.7%) | 2 (16.7%) | 1 (8.3%) | 0 (0.0%) | 1 (8.3%) |
Other techniquesb | 28 (16.7%) | 2 (7.1%) | 10 (35.7%) | 8 (28.6%) | 7 (25.0%) | 1 (3.6%) | 0 (0.0%) |
Total (accounting for duplicates) | 169 | 37 (21.9%) | 56 (33.1%) | 42 (24.9%) | 26 (15.4%) | 4 (2.37%) | 4 (2.37%) |
Approaches to validation
Approach to validationb | |||||||
---|---|---|---|---|---|---|---|
Outcome predicted | Total papersa | Validated | Independent data | Leave-P-out | k-fold cross-validation | Randomly selected subset | Otherb |
Complications | 79 (46.7%) | 73 (92.4%) | 5 (6.85%) | 5 (6.85%) | 33 (45.2%) | 30 (41.1%) | 0 (0%) |
Mortality | 70 (41.4%) | 68 (97.1%) | 5 (7.35%) | 3 (4.41%) | 33 (48.5%) | 27 (39.7%) | 0 (0%) |
Length of stay | 18 (10.7%) | 18 (100%) | 3 (16.7%) | 1 (5.56%) | 4 (22.2%) | 10 (55.6%) | 1 (5.6%) |
Health improvement | 17 (10.1%) | 16 (94.1%) | 0 (0%) | 1 (6.25%) | 5 (31.2%) | 10 (56.2%) | 0 (0%) |
Total (accounting for duplicates) | 169 | 161 (94.1%) | 10 (6.2%) | 8 (5%) | 71 (44.1%) | 71 (44.1%) | 1 (0.6%) |
Measures of predictive accuracy reported
Measure of predictive accuracy reporteda | ||||||
---|---|---|---|---|---|---|
Outcome predicted | Total papers | AUC and accuracy/sensitivity/specificity | AUC only | Accuracy/sensitivity/specificity only |
R
2
| Otherb |
Complication | 73 (45.3%) | 24 (32.9%) | 17 (23.3%) | 28 (38.4%) | 4 (5.5%) | |
Mortality | 68 (42.2%) | 16 (23.5%) | 31 (45.6%) | 18 (26.5%) | 3 (4.4%) | |
Length of stay | 18 (11.1%) | 2 (11.8%) | 3 (16.7%) | 5 (27.8%) | 8 (44.4%) | 1 (5.6%) |
Health improvement | 16 (10%) | 1 (6.3%) | 3 (18.8%) | 11 (68.8%) | 1 (6.3%) | |
Total | 161 | 43 (26.7%) | 54 (33.5%) | 62 (38.5%) | 8 (5.0%) | 9 (5.6%) |