1. Track populations through pattern recognition [EL2]. Example: Peer-to-peer bluetooth apps on smartphones may prevent potentially dangerous contacts (not so much AI needed for that). More challenging: use deep NNs to recognise faces or gaits of persons and their contacts in videos. Detect mass behavior and predict outbreaks and other consequences to build early warning systems (compare Covid-19 forecasting challenge [KAG1]). This may be harder in countries with strict privacy laws [GEO] [DEC]. Identify groups at risk and predict results of therapies. Predict future demand for limited resources (ventilators, doctors) to optimize logistics [EL1]. Sequence virus genomes and detect their cities of origin [EL1]; predict where similar genomes will show up next. Build causal models of the spread of the disease [EL2]. Use NNs to obtain improved epidemiological models from data [DER].
2. Observe single patients. Teach NNs to monitor bio signals, heart rates (e.g., from smart watches), breathing [CHO], coughs [IMR] [EL2], other signals, e.g., [HYL]. Detect & predict asymptomatic cases in time. Analyze X-ray [MAG] [EL2] and other types of images; diagnose pathologies. (The first medical imaging contest won by NNs dates back to 2012 [MED] [TOP1] [GPUCNN5] when compute was almost 100 times more expensive than today.)
3. Partially automate drug design [EL1] and use AI to advance the field of immunology. Find molecules that dock on the (few) proteins of the simple virus to inhibit its activity (like antibodies). E.g., predict folding of proteins to find docking stations. Already 13 years ago when compute was almost 1000 times more expensive than today, Long Short-Term Memory (LSTM) excelled at protein folding prediction [HO1]. See also Google DeepMind's recent computational predictions of protein structures associated with Covid-19 [DMCO].