We collected demographics, validated vital signs, respiratory support settings, medications, and medical history on 433 patients under 24 months of age placed on HFNC in the Johns Hopkins Children’s Center Pediatric Intensive Care Unit from January 2019 through October 2020. Tree-based machine learning algorithms were trained to predict flow rate escalation at lead times varying from 1 to 12 hours. A long short-term memory (LSTM) neural network was trained to forecast future HFNC flow rates based on multivariate sequence data collected from a patient’s electronic health record.

Team Members

[foreach 357]

• [if 397 not_equal=””][/if 397][395]

[/foreach 357]

Project Mentors, Sponsors, and Partners

• Jules Bergmann

• Anthony Sochet

• James Fackler

• Kirby Gong

• Indranuj Gangan

• Raimond L Winslow

• Joseph L Greenstein

Course Faculty

[foreach 429]

• [if 433 not_equal=””][/if 433][431]

[/foreach 429]

Project Links

• Project Poster Image

Additional Project Information

Video Transcript:

Hello, our project is Using Advanced Machine Learning Models to Predict Flow Rate Escalation for Pediatric Patients on High Flow Nasal Cannula. High flow nasal cannula, or HFNC, is commonly used as non-invasive respiratory support in critically ill children. Clinical scores, such as the ROX index, have been used to predict HFNC failure, but they focus only on escalation to mechanical ventilation. However, in response to a patient’s failure on HFNC, depending on the severity, a clinician may also have the ability to increase the oxygen flow rate on the device itself and prevent mechanical ventilation altogether.

This brings us to our objective: to evaluate the abilities of tree-based and neural network machine learning algorithms to predict HFNC flow escalation and to forecast future flow rates.

On the left side of our results section, you can see that our top-performing gradient-boosting model achieved an AUROC of 0.810 for 1 hour before a patient’s flow rate escalation. In the middle of the results section, you can see that our model outperforms the ROX index at short and long lead times. Lastly, on the right side of our results section, you can see the error distribution of our LSTM in forecasting HFNC flow rates for both ventilated and non-ventilated patients.

This leads us to the discussion. Why was the ROX index inaccurate in predicting flow rate escalation? The explanation could lay in the fact that this score relies on the degree of clinical intervention. The same patient could have a different ROX score depending on how actively the clinician lowers the FiO2 upon stabilization. Knowing this, we incorporated two features into our model: the number of times FiO2 was changed and how often the patient’s SpO2 saturated without the clinician lowering the FiO2.

In conclusion, our gradient-boosting models outperform the ROX index in predicting increased flow rate. Lastly, our LSTM has the potential to forecast future flow rates based on a patient’s existing electronic health record and real-time physiologic time series data.