By now, we’re all aware that Covid-19 symptoms vary widely from person to person, though doctors still aren’t sure why. Apart from age, what causes some people to have a mild case while others fight for their lives? It may be a while before scientists can conclusively answer that question. In the meantime, what if deep learning could predictively sort out the Mild Melindas from the Severe Salvadors—before the illness became critical?

General Risk Factors

So what are some of the determining factors that indicate a propensity for a more severe illness? A wide-ranging study conducted recently by the University of Barcelona examined populations in 126 countries and determined a significant strong or moderate positive correlation for the following health-related factors in individuals:

  • Obesity in males
  • Cancer across all populations, especially lung cancer
  • Type-1 diabetes in children

The study also found a significant positive correlation for the following behavioral factors:

  • Alcohol consumption
  • Smoking

Finally, the study found a strong negative correlation for Vitamin D blood serum levels; those with higher levels of vitamin D were less likely to contract Covid.

None of these factors should come as a surprise. Healthy living reduces risk.

Deep Learning Identifies Low- and High-Risk Symptom Clusters

While it is helpful to get a general picture, knowing this information is of limited help in trying to predict who will become very ill once infection has occurred. Fortunately, Covid symptoms themselves offer some clues that are being successfully analyzed by artificial intelligence. A collaboration of biomedical engineers and geneticists from King’s College in London, epidemiologists at Massachusetts General Hospital, and clinicians from Sweden used deep learning to identify clusters of symptoms to predict which Covid patients would experience severe cases requiring respiratory support.

The training dataset was obtained from responses given by users of a smartphone app. Of the 1,653 participants, 383 reported at least one hospital visit and 107 reported needing respiratory support (ventilation or supplementary oxygen). Data was compared against an independent replication sample of 1,047 participants, of which 207 reported a hospital visit and 59 received respiratory support.

The researchers ran the data through an unsupervised machine learning algorithm that clustered six specific groups of symptoms together. One symptom that all six groups consistently reported was headache, but other symptom severity and duration varied widely among the groups. The six clusters of symptoms that the algorithm created in the training set are as follows, with most common symptoms listed earlier and those with more extensive duration noted:

  • Cluster 1 — This group reported sore throat, persistent cough, loss of smell, unusual muscle pains, and chest pain. 1.5% of patients in this cluster required respiratory support, and 16% made at least one visit in the hospital. This was the most common symptom cluster.
  • Cluster 2 — Persistent fever, cough, sore throat, hoarse voice, skipped meals, and a brief loss of smell characterized this group. 4.4% required respiratory support, and 17.5% visited the hospital.
  • Cluster 3 — Loss of smell, diarrhea, many skipped meals, chest pain, and a short-lived sore throat were the main symptoms in this group. 3.7% required respiratory support, and 23.6% made at least one trip to the hospital.
  • Cluster 4 — Persistent cough, a brief loss of smell, hoarse voice for days, recurring instances of chest pain and fatigue, fever, and fatigue, were experienced by those in this group. 8.6% required respiratory support and 24.6% went to the hospital.
  • Cluster 5 — This cluster experienced intense unusual muscle pains, sore throat, many skipped meals, persistent cough, hoarse voice, frequent episodes of confusion, prolonged fatigue, fever, and chest pain. Out of this group, 9.9% required respiratory support, and 27.2% ended up in the hospital.
  • Cluster 6 — Cluster 6 had the greatest number of symptoms, with a high prevalence of unusual muscle pains, long-lasting abdominal pain, and skipped meals. Severe diarrhea, persistent cough sore throat, chest pain, and frequent shortness of breath accompanied by fever and severe fatigue also characterized this group. Nearly 20% required respiratory support, and almost half (45.5%) went to the hospital at least once.


Summary and Conclusion

People in the “milder” clusters (1 and 2) tended to be younger and have a lower BMI. People in Cluster 3 were similar demographically to Clusters 1 and 2 but ended up in the hospital somewhat more frequently due to ongoing gastrointestinal issues. The severe presentation and duration of confusion, fatigue, and skipped meals were hallmarks of Clusters 5 and 6, and those groups tended to be older and frailer.

These data were replicated with a “modest” margin of error in the independent sample cases. A noted weak point of the study was that the data was inherently limited in being self-reported, and researchers acknowledged that some individuals may have become too unwell to accurately report symptoms. Ideally, a larger sample size would create a more robust data set.

Still, having an automated clinical tool that could red-flag patients likely to need high-level care by analyzing their specific combinations of early-onset symptoms could go a long way in helping doctors strategically apply limited resources in this pandemic, consequently providing better care for those at greatest risk.

Author: Erika Greelish