Improved ICU mortality prediction based on SOFA scores and gastrointestinal parameters

Discussion

There is an ongoing effort to improve prediction models for patient outcome in the ICU. In this study we tested the efficacy of using a patient’s latest SOFA scores to represent the change in condition throughout ICU stay for the purpose of predicting ICU mortality. We first examined the ability of the SOFA score to predict mortality on the using the data from our ICU. The need to use sub-scores dictates larger input vectors[ 9 ], thus in this work we examined new ways to achieve this level of accuracy with more compact inputs. Using several machine learning algorithms showed good performance of the SOFA score with an AUC mostly above 0.9. We then assessed several ensemble methods and found the combination of logistic and linear regression to slightly improve prediction. Furthermore, since so many models and methodologies were used, examining the different models we observed a range of performance in accuracy, showing a relatively tight interval between 0.8875 and 0.9113. This narrow interval, despite using four different algorithms, ensembles and input combinations, indicates solid results where accuracy is not expected to decline drastically when further tested on new data, possibly from mixed center populations (i.e., patients from other hospitals/countries). The next step was to incorporate a GI failure score with the SOFA score to further improve prediction accuracy. We used descriptive decision trees to discover GI parameters that may be able to reduce prediction error of classifiers based solely on SOFA. In the aforementioned study by Reintam et al. [ 17 ], a GI dysfunction score was developed in an effort to further improve the performance of the SOFA score; however, the results were equivocal [ 16 ]: although the number of GI symptoms was significantly higher in non-survivors, no symptom could be used as an independent predictor for mortality. Furthermore, the incorporation of the combination of SOFA and GI failure scale to this new heterogenic population failed to improve performance. The final conclusion drawn from these past studies was that a new approach to the problem was required.

It seems that a few obstacles prohibit the GI system's incorporation into severity scoring systems, including the wide diversity of gastrointestinal disorder clinical manifestations in the ICU [ 25 ], a lack of an accepted definition for GI failure [ 26 ], lacking validation of laboratory markers, mainly citrulline [ 27 ], and the scarcity of strong-level evidence. Feeding intolerance, an important manifestation and defining factor for GI failure, is by itself not yet well defined [ 28 ], as it may be based solely on GRV measurements, amount of enteral nutrition delivered or GI symptom lists. Understanding of the intricate interrelation between acute GI dysfunction, feeding intolerance and intraabdominal hypertension and their wide areas of overlap is still evolving [ 29 ].

We devised a completely new approach for the incorporation of the GI abnormalities into prognostic methods. Our machine learning prediction model combines integrated gastrointestinal disturbances with well-established organ failure severity score. The model significantly improved the prediction capabilities of the standard SOFA score. Moreover, the model analyzes the dynamics of change in these parameters over time, making it a dynamic score (i.e., adding the important element of time). The time series approach allows for a significant improvement in mortality risk prediction compared to a single SOFA score reading. Our research shows that our approach allows the design of a prediction model with improved prediction accuracy of ICU mortality risk, potentially advancing towards the addition of GI component into the SOFA score, thus improving its predictive abilities.

Conclusions

Our models of data analysis yielded strong evidence for the accuracy of the SOFA-based scoring system. When incorporating the time element by looking at three consecutive SOFA scores and adding a seventh we demonstrated a yet more accurate predictive ability of the model. We believe it represents a step towards a call for the inclusion of the GI system in SOFA-based scoring systems and helps bridge the evidence gap in this field.

Abbreviations

Funding Statement

The authors received no specific funding for this work.

Data Availability

The public sharing of the data underlying this study is restricted as per the policy of the data guardian, Clalit Health Services, as the data contain sensitive patient information. Although the authors cannot make their study’s data publicly available at the time of publication, all authors commit to make the data underlying the findings described in this study fully available without restriction to those who request the data, in compliance with the PLOS Data Availability policy. For data sets involving personally identifiable information or other sensitive data, data sharing is contingent on the data being handled appropriately by the data requester and in accordance with all applicable local requirements. Data access queries may be directed to Dr. Itai Bendavid ( moc.allaw@dbti ) or Prof. Pierre Singer ( li.gro.tilalc@regnisp ). Per the requirements of Clalit Health Services, data requesters can only access the hospital's dataset locally (i.e, physically in our hospital) under the supervision of Dr. Bendavid and Prof. Singer, and the data cannot exported, even in anonymized form.

References