Post by dimitrytran on Mar 21, 2020 10:49:57 GMT
Hi everyone, by way of introduction, I am a co-founder of Harrison.ai. a company based in Sydney. Please find below a short summary of the idea. I can circulate a more detailed 2-pager proposal if there is interest.
BACKGROUND
Our first product currently in wide clinical use is in the field of IVF, where we build AI to predict the best embryo that resulting in pregnancy - NewScientist Article. We are currently working in partnership with the world's second largest radiology company, I-MED, with over 6+ million care episodes a year, to build a comprehensive AI in Chest X-Ray with 100+ findings. Many of these findings are relevant to findings of COVID (e.g. airspace opacity and the differences between focal, diffuse, interstitial and location e.g. 21 critically ill patients in Washington - JAMA article.)
PROBLEM STATEMENT
We noted that approximately 21% of inpatients go on to developed ARDS and require intubation. Conversely this means that if you can predict who’s going to develop ARDS, you could discharge up to 79% of your population or move them to a lower acuity care setting.
Lower numbers of inpatients will reduce burden on hospital staff and reduce inpatient transmission, and help plan logistics and resource allocation. Serial CXR could be a potential outpatient monitoring tool if we can accurate predict the clinical course.
PROPOSAL
The proposal is to create an AI Tool to make triage predictions relevant to COVID-19 cases and implement it in all jurisdictions where the input parameters are available in digital form.
The hypothesis is that an AI tool can be created that will provide consistent Triage of COVID-19 cases in 3 settings
1. Inpatient vs outpatient support
Predicting likelihood of patient surviving without escalation of care to NIV or other care that can only be delivered in a hospital (ward level, not ICU care)
Benefit: safely have less patients occupying hospital beds who will not need them
2. Invasive ventilatory support
Predicting likelihood of patient surviving if intubated and given full ICU support
Benefit: safely triage patients onto ventilation with a higher probability of survival, when near maximal ventilation capacity is reached
3. Removal of invasive ventilatory support
Predicting likelihood of patient surviving once intubated and receiving full ICU support
Benefit: safely triage patients off ventilation who have a very low probability of survival, allowing others with higher probability of benefit to access ventilation, once maximal ventilation capacity is exceeded.
REQUIREMETNS
Given the CXR AI is already pre-trained, we think a small dataset from the NHS (500-1000 cases) would be sufficient to develop the prediction model.
Input data required
Outcome data required
Look forward to any feedback from the group on this idea.
I am available at dimitry@harrison.ai and WhatsApp at +61 402 511 469
BACKGROUND
Our first product currently in wide clinical use is in the field of IVF, where we build AI to predict the best embryo that resulting in pregnancy - NewScientist Article. We are currently working in partnership with the world's second largest radiology company, I-MED, with over 6+ million care episodes a year, to build a comprehensive AI in Chest X-Ray with 100+ findings. Many of these findings are relevant to findings of COVID (e.g. airspace opacity and the differences between focal, diffuse, interstitial and location e.g. 21 critically ill patients in Washington - JAMA article.)
PROBLEM STATEMENT
We noted that approximately 21% of inpatients go on to developed ARDS and require intubation. Conversely this means that if you can predict who’s going to develop ARDS, you could discharge up to 79% of your population or move them to a lower acuity care setting.
Lower numbers of inpatients will reduce burden on hospital staff and reduce inpatient transmission, and help plan logistics and resource allocation. Serial CXR could be a potential outpatient monitoring tool if we can accurate predict the clinical course.
PROPOSAL
The proposal is to create an AI Tool to make triage predictions relevant to COVID-19 cases and implement it in all jurisdictions where the input parameters are available in digital form.
The hypothesis is that an AI tool can be created that will provide consistent Triage of COVID-19 cases in 3 settings
1. Inpatient vs outpatient support
Predicting likelihood of patient surviving without escalation of care to NIV or other care that can only be delivered in a hospital (ward level, not ICU care)
Benefit: safely have less patients occupying hospital beds who will not need them
2. Invasive ventilatory support
Predicting likelihood of patient surviving if intubated and given full ICU support
Benefit: safely triage patients onto ventilation with a higher probability of survival, when near maximal ventilation capacity is reached
3. Removal of invasive ventilatory support
Predicting likelihood of patient surviving once intubated and receiving full ICU support
Benefit: safely triage patients off ventilation who have a very low probability of survival, allowing others with higher probability of benefit to access ventilation, once maximal ventilation capacity is exceeded.
REQUIREMETNS
Given the CXR AI is already pre-trained, we think a small dataset from the NHS (500-1000 cases) would be sufficient to develop the prediction model.
Input data required
- Patient parameters, such as age, co-morbidities and pre-existing diseases, and real time measures of severity of illness (WCC, Fibrinogen, LFTs, CRP etc.)
- Acute measures of progression (RR, SpO2, Ventilatory support such as PEEP, MAwP, ECMO, A-a Gradient, P/F ratio, SOFA score)
- Imaging Progression – CXR and Chest CT
(if available)
Outcome data required
- Setting
- Need for hospital based care (NIV, Nursing dependency)
- Survival of patients offered & commenced on invasive ventilation in ICU
- Survival of patients on invasive ventilation in ICU (over time, with progression)
In the longer term, survival may have to be replaced with longer term measures of functional status post hospital discharge.
Look forward to any feedback from the group on this idea.
I am available at dimitry@harrison.ai and WhatsApp at +61 402 511 469