Predicting ICU Need

During the Covid-19 Pandemic

Charles C Barnes

QuickStart, Inc.

July 30, 2024

The Problem

We are facing overcapacity of the hospital system globally from the COVID-19 pandemic.

• Decreased capacity to treat people with severe cases of covid in the intensive care unit
• Decreased capacity to treat people without COVID who need the ICU

Not everyone admitted to the hospital for COVID requires the ICU

A Solution

Can we predict whether a person will need ICU treatment?

• Increase recovery/stability potential
• Decrease time to recovery/stability¹
• Effective resource use for people who need it most
• Bed space availability

1. Mitigating viral load early on improves recovery (Sanche et al. 2022)

Our Dataset

Across 385 people treated for COVID:

• demographics
• disease history
• blood test results
• vital signs
• time window of admittance to ICU

What was the ICU need?

163 people were moved to the ICU and 190 people never needed the ICU.

People moved to the ICU

The people who needed the ICU¹:

• were biased by gender
• were older than 65 by ~30% increased representation
• had comorbidities by ~15% increased representation

1. Significant by Fisher’s exact test (p<0.05)

PCR¹ and predictive measures

• Na
• Ca
• Resp

People with different ICU needs overlap substantially.

1. A test detecting coronavirus. Viral load correlates with disease severity (Sanche et al. 2022)

Modeling decisions

Filled missing values using neighboring values

Target feature engineering:

• identify people moved to ICU and recode as ‘target’
• drop data entries during and after ICU admission

Feature selection

• to remove highly correlated features

Random undersampling ‘No’ ICU need

• to match amount of ‘Yes’ values.

70:15:15 train-test-validation split

• 644:138:138 samples representing people¹

Logistic regression

Decision tree

Random forest

Best features

• model accuracy and variance among models

Model tuning

• scoring to minimize the difference between predicted and actual ‘target’ values

1. People admitted to the hospital for COVID treatment

Final model performance

Random forest model:

• 50 features

ICU need: Yes

• 89% correct
• recalled 66 of 70 people

ICU need: No

• 94% correct
• recalled 62 of 70 people

Recommendation

I recommend that the model be used to supplement medical expertise and discretion when identifying people who need more monitoring and may ultimately be moved to the ICU.

Changes from implementing the model for triage:

• What is the outcome for people treated in the ICU?
• What is the ICU stay duration?
• What is the total hospital stay duration?
• Hospital resources?
• Ability to meet other ICU needs?

Limitations

Model (what this model addresses):

• Whether people will need the ICU or not
• NOT: when will they require the ICU?
  • Regressor for time moved to the ICU

Dataset (what we can ask of the data):

• Do results generalize to people globally?
• When were people moved to the ICU¹?
  • “above-12” can include 13 or 24 hours after being admitted
• Could we predict recovery time too?²

1. Was this delay due to symptom worsening or no beds available?

2. We now know that recovery is dependent upon viral load and mitigating it as soon as possible (Sanche et al. 2022).

The model and data fits our primary need to quickly identify whether or not people will need the ICU.

Thank you

For the work you do to meet this challenge head on with a face mask!

References

Sanche, Steven, Tyler Cassidy, Pinghan Chu, Alan S. Perelson, Ruy M. Ribeiro, and Ruian Ke. 2022. “A simple model of COVID-19 explains disease severity and the effect of treatments.” Scientific Reports 12 (1): 1–14. https://doi.org/10.1038/s41598-022-18244-2.