Retrospective studies demonstrate that machine learning models can accurately predict sepsis and septic shock onset. Prospective clinical studies at the bedside are needed to assess their effect on patient-relevant outcomes. |
Introduction
Methods
Search strategy
Eligibility criteria and study selection
Quality of evidence and risk of bias
Performance metric and meta-analysis
Results
Study selection
Study characteristics
Paper | Design | Target condition | Patient encounters | Machine learning model | Comparators | |
---|---|---|---|---|---|---|
Validation | ||||||
ED | Brown et al. | Prospective validation | Severe sepsis and Septic shock | 93,773 (15 months) | Cut05 Primary outcome Sensitivity: 0.764 False positive rate: 0.47 Secondary outcome AUC: 0.859 | Nurse triage Primary outcome Sensitivity: 0.543 False positive rate: 0.31 Secondary outcome AUC: 0.756 |
SIRS Primary outcome Sensitivity: 0.216 False positive rate: 0.004 Secondary outcome AUC: 0.606 | ||||||
In-hospital | Thiel et al. | Prospective validation | Septic shock | 27,674 (24 months) | RPARTa 2006 Primary outcome Misclassification rate: 8.4% | None |
RPARTa 2007 Primary outcome Misclassification rate: 8.8% |
Paper | Design | Target condition | Patient encounters | Machine learning group | Control group | |
---|---|---|---|---|---|---|
Interventional | ||||||
In-hospital | Giannini et al. | Pre-post implementation | Severe sepsis and septic shock | 54,464 (6 pre-months, 1 post-month) | EWS 2.0 Primary/secondary outcome Hospital LOS: 9 days Time to ICU transfer after alert: 8 he In-hospital mortality: 10.3% | Unclear Primary/secondary outcome Hospital LOS: 9 days Time to ICU transfer after alert: 16 he In-hospital mortality: 10.6% |
McCoy et al. | Pre-post implementationb | Severe sepsis | 611 (3 pre-months, 2 post-months) | Linear model (Insight) Primary outcome In-hospital mortality: 2.94% Secondary outcome Hospital LOS: 2.92 days Readmission rate: 7.84% | Manual nurse scoringc Primary outcome In-hospital mortality: 7.37% Secondary outcome Hospital LOS: 3.35 days Readmission rate: 46.19% | |
ICU | Shimabukuro et al. | RCT | Severe sepsis | 142 (3 months) | Elastic net reg.d (Insight) Primary outcome Hospital LOS: 10.3 dayse Secondary outcome ICU LOS: 6.3 dayse In-hospital mortality: 8.96%e | SIRS detector Primary outcome Hospital LOS: 13.0 dayse Secondary outcome ICU LOS: 8.4 dayse In-hospital mortality: 21.3%e |
EDa 4 papers 22 models | In-hospitala 7 papers 43 models | ICUa 15 papers 52 models | ||||
---|---|---|---|---|---|---|
Absolute | Proportion | Absolute | Proportion | Absolute | Proportion | |
Per paper | ||||||
Prospective design | 1 | 0.25 | 2 | 0.29 | 1 | 0.07 |
Privacy statement | 0 | 0.00 | 3 | 0.43 | 5 | 0.33 |
MIMICb | – | – | – | – | 9 | 0.60 |
Description of patients | 4 | 1.00 | 2 | 0.29 | 5 | 0.33 |
Inclusion criteria | 3 | 0.75 | 4 | 0.57 | 12 | 0.80 |
Country—USA | 4 | 1.00 | 7 | 1.00 | 15 | 1.00 |
Per model | ||||||
Target condition | ||||||
Sepsis | 20 | 0.91 | 10 | 0.23 | 37 | 0.71 |
Severe sepsis | 0 | 0.00 | 1 | 0.02 | 12 | 0.23 |
Severe sepsis & septic shock | 2 | 0.09 | 1 | 0.02 | 0 | 0.00 |
Septic shock | 0 | 0.00 | 31 | 0.72 | 3 | 0.06 |
Components of target condition definition | ||||||
ICD | 20 | 0.91 | 32 | 0.74 | 17 | 0.33 |
SIRS | 0 | 0.00 | 4 | 0.09 | 19 | 0.37 |
SOFA | 0 | 0.00 | 1 | 0.02 | 21 | 0.40 |
Data split design | ||||||
Train-(validate)-test | 20 | 0.91 | 15 | 0.35 | 21 | 0.40 |
Cross-validation | 0 | 0.00 | 25 | 0.58 | 28 | 0.54 |
Data granularity | ||||||
1-hourly values | – | – | 6 | 0.14 | 30 | 0.58 |
> 1/hourly values | – | – | 0 | 0.00 | 18 | 0.35 |
Not described | – | – | 31 | 0.72 | 4 | 0.08 |
Missing values strategies | ||||||
Feedforward | 0 | 0.00 | 8 | 0.19 | 14 | 0.27 |
Mean imputation | 0 | 0.00 | 9 | 0.21 | 12 | 0.23 |
Zero imputation | 0 | 0.00 | 16 | 0.37 | 0 | 0.00 |
Nearest neighbor | 0 | 0.00 | 0 | 0.00 | 16 | 0.31 |
Physiological imputation | 13 | 0.59 | 0 | 0.00 | 0 | 0.00 |
Otherc | 7 | 0.32 | 3 | 0.07 | 2 | 0.04 |
Not described | 2 | 0.09 | 7 | 0.16 | 8 | 0.15 |
Model | ||||||
Generalized linear model | 3 | 0.14 | 6 | 0.14 | 15 | 0.29 |
Naïve Bayes | 11 | 0.50 | 3 | 0.07 | 0 | 0.00 |
Ensemble methods | 4 | 0.18 | 9 | 0.21 | 7 | 0.13 |
Proportional hazard | 0 | 0.00 | 0 | 0.00 | 9 | 0.17 |
Decision tree | 0 | 0.00 | 9 | 0.21 | 0 | 0.00 |
Support vector machines | 4 | 0.18 | 3 | 0.07 | 11 | 0.21 |
Neural network | 0 | 0.00 | 8 | 0.19 | 6 | 0.12 |
Long short-term memory (LSTM) | 0 | 0.00 | 5 | 0.12 | 4 | 0.08 |
Quality of evidence and risk of bias
Paper | Setting | Risk of bias | |||||
---|---|---|---|---|---|---|---|
Patient selection | Index test | Reference standard | Flow and timing | Data management | |||
ED | Horng et al. [35] | Sepsis | ? | ||||
Haug et al. [62] | Sepsis | ? | |||||
Delahanty et al. [63] | Sepsis | ? | |||||
Brown et al. [28] | Severe sepsis and septic shock | ? | |||||
In-hospital | Khojandi et al. [64] | Sepsis | ? | ||||
Futoma et al. [65] | Sepsis | ? | |||||
McCoy et al. [31] | Severe sepsis, Sepsis | ? | |||||
Lin et al. [66] | Septic shock | ? | |||||
Khoshnevisan et al. [14] | Septic shock | ? | |||||
Thiel et al. [29] | Septic shock | ? | |||||
Giannini et al. [30] | Severe sepsis and septic shock | ? | |||||
ICU | Wang et al. [57] | Sepsis | ? | ||||
Shashikumar II et al. [67] | Sepsis | ? | |||||
Shashikumar I et al. [68] | Sepsis | ? | |||||
Scherpf et al. [59] | Sepsis | ? | |||||
Desautels et al. [54] | Sepsis | ? | |||||
Nemati et al. [55] | Sepsis | ? | |||||
Calvert II et al. [52] | Sepsis | ? | |||||
Kam et al. [53] | Sepsis | ? | |||||
Van Wyk I et al. [69] | Sepsis | ? | |||||
Van Wyk II et al. [70] | Sepsis | ? | |||||
Moss et al. [36] | Severe sepsis | ? | |||||
Guillén et al. [58] | Severe sepsis | ? | |||||
Shimabukuro et al. [32] | Severe sepsis | ? | |||||
Henry et al. [56] | Septic shock | ? | |||||
Calvert I et al. [33] | Septic shock | ? | |||||
ED/In-hospital/ICU | Barton et al. [27] | Sepsis | ? | ||||
Mao et al. [71] | Sepsis, severe sepsis, septic shock | ? |
Study Characteristics | Quality Assessment | Outcome | |||||||
---|---|---|---|---|---|---|---|---|---|
No of studies | Design | Limitations (Unclear risk of bias studies/total) | Indirectness of patients, settingb | Indirectness of outcome | Inconsistencyc | Imprecision | AUROC high risk of bias/unclear risk of bias | Quality of evidence | |
ED | Sepsis | ||||||||
3 studies (3.270.608 patients) | Cohort studies | High risk of bias (2/3) | None | Serious indirectness—differences in outcome definition | Not available | None | 0.95–0.97/0.65–0.97 | ⊕ ⊕ ⊙ ⊙ Low | |
In-hospital | Septic shock | ||||||||
2 studies (51,540 patients) | Cohort studies | High risk of bias (0/2) | None | Serious indirectness—differences in outcome definition | Not available | None | 0.86–0.94 | ⊕ ⊕ ⊙ ⊙ Low | |
ICU | Sepsis | ||||||||
8 studies (125.162 patients) | Cohort studies | High risk of bias (2/8) | None | Serious indirectness—differences in outcome definition | Not available | None | 0.70–0.99/0.81–0.88 | ⊕ ⊕ ⊙ ⊙ Low | |
Severe sepsis | |||||||||
3 studies (6.647 patients) | Cohort studies | High risk of bias (0/3) | None | Serious indirectness—differences in outcome definition | Not available | None | 0.68–0.95 | ⊕ ⊕ ⊙ ⊙ Low | |
Septic shock | |||||||||
2 studies (16.234 patientsa) | Cohort studies | High risk of bias (1/2) | None | Serious indirectness—differences in outcome definition | Not available | None | 0.89–0.96/0.83–0.83 | ⊕ ⊕ ⊙ ⊙ Low |
Meta-analysis
Variables | Univariate analysis | Multivariate analysis | ||||
---|---|---|---|---|---|---|
Coeff | SE | p value | Coeff | SE | p value | |
Temperature as feature | 0.788 | 0.239 | 0.002 | 0.812 | 0.218 | 0.000 |
Lab values as feature | 0.835 | 0.311 | 0.008 | 0.842 | 0.291 | 0.003 |
Type of model (ref. = EM) | 0.018 | 0.020 | ||||
Generalized linear model | − 0.211 | 0.251 | − 0.211 | 0.231 | ||
Naïve Bayes | − 0.651 | 0.312 | − 0.682 | 0.291 | ||
Neural network | 0.344 | 0.300 | 0.172 | 0.278 | ||
Proportional hazard | − 0.464 | 0.851 | − 0.506 | 0.673 | ||
Support vector machines | − 0.168 | 0.256 | − 0.161 | 0.241 | ||
Decision trees | − 1.013 | 0.419 | − 1.088 | 0.399 | ||
Target condition defined as Seymour (Sepsis-3) | − 1.039 | 0.459 | 0.025 | |||
Target condition definition contains SOFA | − 0.935 | 0.438 | 0.033 | |||
Respiratory rate as feature | 0.672 | 0.250 | 0.008 | |||
Heart rate as feature | 0.680 | 0.327 | 0.037 | |||
Arterial blood gas as feature | 0.802 | 0.313 | 0.011 |