README.md

---
license: mit
pipeline_tag: question-answering
widget:
- text: What is the delay between illness onset and infection?
  context: 'Epidemiological research priorities for public health control of the ongoing
    global novel coronavirus (2019-nCoV) outbreak https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7029449/
    SHA: 90de2d957e1960b948b8c38c9877f9eca983f9eb Authors: Cowling, Benjamin J; Leung,
    Gabriel M Date: 2020-02-13 DOI: 10.2807/1560-7917.es.2020.25.6.2000110 License:
    cc-by Abstract: Infections with 2019-nCoV can spread from person to person, and
    in the earliest phase of the outbreak the basic reproductive number was estimated
    to be around 2.2, assuming a mean serial interval of 7.5 days [2]. The serial
    interval was not precisely estimated, and a potentially shorter mean serial interval
    would have corresponded to a slightly lower basic reproductive number. Control
    measures and changes in population behaviour later in January should have reduced
    the effective reproductive number. However, it is too early to estimate whether
    the effective reproductive number has been reduced to below the critical threshold
    of 1 because cases currently being detected and reported would have mostly been
    infected in mid- to late-January. Average delays between infection and illness
    onset have been estimated at around 5–6 days, with an upper limit of around 11-14
    days [2,5], and delays from illness onset to laboratory confirmation added a further
    10 days on average [2]. Text: It is now 6 weeks since Chinese health authorities
    announced the discovery of a novel coronavirus (2019-nCoV) [1] causing a cluster
    of pneumonia cases in Wuhan, the major transport hub of central China. The earliest
    human infections had occurred by early December 2019, and a large wet market in
    central Wuhan was linked to most, but not all, of the initial cases [2] . While
    evidence from the initial outbreak investigations seemed to suggest that 2019-nCoV
    could not easily spread between humans [3] , it is now very clear that infections
    have been spreading from person to person [2] . We recently estimated that more
    than 75,000 infections may have occurred in Wuhan as at 25 January 2020 [4] ,
    and increasing numbers of infections continue to be detected in other cities in
    mainland China and around the world. A number of important characteristics of
    2019-nCoV infection have already been identified, but in order to calibrate public
    health responses we need improved information on transmission dynamics, severity
    of the disease, immunity, and the impact of control and mitigation measures that
    have been applied to date. Infections with 2019-nCoV can spread from person to
    person, and in the earliest phase of the outbreak the basic reproductive number
    was estimated to be around 2.2, assuming a mean serial interval of 7.5 days [2]
    . The serial interval was not precisely estimated, and a potentially shorter mean
    serial interval would have corresponded to a slightly lower basic reproductive
    number. Control measures and changes in population behaviour later in January
    should have reduced the effective reproductive number. However, it is too early
    to estimate whether the effective reproductive number has been reduced to below
    the critical threshold of 1 because cases currently being detected and reported
    would have mostly been infected in mid-to late-January. Average delays between
    infection and illness onset have been estimated at around 5-6 days, with an upper
    limit of around 11-14 days [2, 5] , and delays from illness onset to laboratory
    confirmation added a further 10 days on average [2] . Chains of transmission have
    now been reported in a number of locations outside of mainland China. Within the
    coming days or weeks it will become clear whether sustained local transmission
    has been occurring in other cities outside of Hubei province in China, or in other
    countries. If sustained transmission does occur in other locations, it would be
    valuable to determine whether there is variation in transmissibility by location,
    for example because of different behaviours or control measures, or because of
    different environmental conditions. To address the latter, virus survival studies
    can be done in the laboratory to confirm whether there are preferred ranges of
    temperature or humidity for 2019-nCoV transmission to occur. In an analysis of
    the first 425 confirmed cases of infection, 73% of cases with illness onset between
    12 and 22 January reported no exposure to either a wet market or another person
    with symptoms of a respiratory illness [2] . The lack of reported exposure to
    another ill person could be attributed to lack of awareness or recall bias, but
    China''s health minister publicly warned that pre-symptomatic transmission could
    be occurring [6] . Determining the extent to which asymptomatic or pre-symptomatic
    transmission might be occurring is an urgent priority, because it has direct implications
    for public health and hospital infection control. Data on viral shedding dynamics
    could help in assessing duration of infectiousness. For severe acute respiratory
    syndrome-related coronavirus (SARS-CoV), infectivity peaked at around 10 days
    after illness onset [7] , consistent with the peak in viral load at around that
    time [8] . This allowed control of the SARS epidemic through prompt detection
    of cases and strict isolation. For influenza virus infections, virus shedding
    is highest on the day of illness onset and relatively higher from shortly before
    symptom onset until a few days after onset [9] . To date, transmission patterns
    of 2019-nCoV appear more similar to influenza, with contagiousness occurring around
    the time of symptom onset, rather than SARS. Transmission of respiratory viruses
    generally happens through large respiratory droplets, but some respiratory viruses
    can spread through fine particle aerosols [10] , and indirect transmission via
    fomites can also play a role. Coronaviruses can also infect the human gastrointestinal
    tract [11, 12] , and faecal-oral transmission might also play a role in this instance.
    The SARS-CoV superspreading event at Amoy Gardens where more than 300 cases were
    infected was attributed to faecal-oral, then airborne, spread through pressure
    differentials between contaminated effluent pipes, bathroom floor drains and flushing
    toilets [13] . The first large identifiable superspreading event during the present
    2019-nCoV outbreak has apparently taken place on the Diamond Princess cruise liner
    quarantined off the coast of Yokohama, Japan, with at least 130 passengers tested
    positive for 2019-nCoV as at 10 February 2020 [14] . Identifying which modes are
    important for 2019-nCoV transmission would inform the importance of personal protective
    measures such as face masks (and specifically which types) and hand hygiene. The
    first human infections were identified through a surveillance system for pneumonia
    of unknown aetiology, and all of the earliest infections therefore had Modelling
    studies incorporating healthcare capacity and processes pneumonia. It is well
    established that some infections can be severe, particularly in older adults with
    underlying medical conditions [15, 16] , but based on the generally mild clinical
    presentation of 2019-nCoV cases detected outside China, it appears that there
    could be many more mild infections than severe infections. Determining the spectrum
    of clinical manifestations of 2019-nCoV infections is perhaps the most urgent
    research priority, because it determines the strength of public health response
    required. If the seriousness of infection is similar to the 1918/19 Spanish influenza,
    and therefore at the upper end of severity scales in influenza pandemic plans,
    the same responses would be warranted for 2019-nCoV as for the most severe influenza
    pandemics. If, however, the seriousness of infection is similar to seasonal influenza,
    especially during milder seasons, mitigation measures could be tuned accordingly.
    Beyond a robust assessment of overall severity, it is also important to determine
    high risk groups. Infections would likely be more severe in older adults, obese
    individuals or those with underlying medical conditions, but there have not yet
    been reports of severity of infections in pregnant women, and very few cases have
    been reported in children [2] . Those under 18 years are a critical group to study
    in order to tease out the relative roles of susceptibility vs severity as possible
    underlying causes for the very rare recorded instances of infection in this age
    group. Are children protected from infection or do they not fall ill after infection?
    If they are naturally immune, which is unlikely, we should understand why; otherwise,
    even if they do not show symptoms, it is important to know if they shed the virus.
    Obviously, the question about virus shedding of those being infected but asymptomatic
    leads to the crucial question of infectivity. Answers to these questions are especially
    pertinent as basis for decisions on school closure as a social distancing intervention,
    which can be hugely disruptive not only for students but also because of its knock-on
    effect for child care and parental duties. Very few children have been confirmed
    2019-nCoV cases so far but that does not necessarily mean that they are less susceptible
    or that they could not be latent carriers. Serosurveys in affected locations could
    inform this, in addition to truly assessing the clinical severity spectrum. Another
    question on susceptibility is regarding whether 2019-nCoV infection confers neutralising
    immunity, usually but not always, indicated by the presence of neutralising antibodies
    in convalescent sera. Some experts already questioned whether the 2019-nCoV may
    behave similarly to MERS-CoV in cases exhibiting mild symptoms without eliciting
    neutralising antibodies [17] . A separate question pertains to the possibility
    of antibody-dependent enhancement of infection or of disease [18, 19] . If either
    of these were to be relevant, the transmission dynamics could become more complex.
    A wide range of control measures can be considered to contain or mitigate an emerging
    infection such as 2019-nCoV. Internationally, the past week has seen an increasing
    number of countries issue travel advisories or outright entry bans on persons
    from Hubei province or China as a whole, as well as substantial cuts in flights
    to and from affected areas out of commercial considerations. Evaluation of these
    mobility restrictions can confirm their potential effectiveness in delaying local
    epidemics [20] , and can also inform when as well as how to lift these restrictions.
    If and when local transmission begins in a particular location, a variety of community
    mitigation measures can be implemented by health authorities to reduce transmission
    and thus reduce the growth rate of an epidemic, reduce the height of the epidemic
    peak and the peak demand on healthcare services, as well as reduce the total number
    of infected persons [21] . A number of social distancing measures have already
    been implemented in Chinese cities in the past few weeks including school and
    workplace closures. It should now be an urgent priority to quantify the effects
    of these measures and specifically whether they can reduce the effective reproductive
    number below 1, because this will guide the response strategies in other locations.
    During the 1918/19 influenza pandemic, cities in the United States, which implemented
    the most aggressive and sustained community measures were the most successful
    ones in mitigating the impact of that pandemic [22] . Similarly to international
    travel interventions, local social distancing measures should be assessed for
    their impact and when they could be safely discontinued, albeit in a coordinated
    and deliberate manner across China such that recrudescence in the epidemic curve
    is minimised. Mobile telephony global positioning system (GPS) data and location
    services data from social media providers such as Baidu and Tencent in China could
    become the first occasion when these data inform outbreak control in real time.
    At the individual level, surgical face masks have often been a particularly visible
    image from affected cities in China. Face masks are essential components of personal
    protective equipment in healthcare settings, and should be recommended for ill
    persons in the community or for those who care for ill persons. However, there
    is now a shortage of supply of masks in China and elsewhere, and debates are ongoing
    about their protective value for uninfected persons in the general community.
    The Table summarises research gaps to guide the public health response identified.
    In conclusion, there are a number of urgent research priorities to inform the
    public health response to the global spread of 2019-nCoV infections. Establishing
    robust estimates of the clinical severity of infections is probably the most pressing,
    because flattening out the surge in hospital admissions would be essential if
    there is a danger of hospitals becoming overwhelmed with patients who require
    inpatient care, not only for those infected with 2019-nCoV but also for urgent
    acute care of patients with other conditions including those scheduled for procedures
    and operations. In addressing the research gaps identified here, there is a need
    for strong collaboration of a competent corps of epidemiological scientists and
    public health workers who have the flexibility to cope with the surge capacity
    required, as well as support from laboratories that can deliver on the ever rising
    demand for diagnostic tests for 2019-nCoV and related sequelae. The readiness
    survey by Reusken et al. in this issue of Eurosurveillance testifies to the rapid
    response and capabilities of laboratories across Europe should the outbreak originating
    in Wuhan reach this continent [23] . In the medium term, we look towards the identification
    of efficacious pharmaceutical agents to prevent and treat what may likely become
    an endemic infection globally. Beyond the first year, one interesting possibility
    in the longer term, perhaps borne of wishful hope, is that after the first few
    epidemic waves, the subsequent endemic re-infections could be of milder severity.
    Particularly if children are being infected and are developing immunity hereafter,
    2019-nCoV could optimistically become the fifth human coronavirus causing the
    common cold. None declared.'
---

# Model Card for Model longluu/Medical-QA-gatortrons-COVID-QA
The model is an extractive Question Answering algorithm that can find an answer to a question by finding a segment in a text. 

## Model Details

### Model Description
The base pretrained model is GatorTronS which was trained on billions of words in various clinical texts (https://huggingface.co/UFNLP/gatortronS). 
Then using the COVID-QA dataset (https://huggingface.co/datasets/covid_qa_deepset), I fine-tuned the model for an extractive Question Answering algorithm that can answer
a question by finding it within a text.

### Model Sources [optional]
The github code associated with the model can be found here: https://github.com/longluu/Medical-QA-extractive.

## Training Details

### Training Data

This dataset contains 2,019 question/answer pairs annotated by volunteer biomedical experts on scientific articles regarding COVID-19 and other medical issues.
The dataset can be found here: https://github.com/deepset-ai/COVID-QA. The preprocessed data can be found here https://huggingface.co/datasets/covid_qa_deepset.

#### Training Hyperparameters

The hyperparameters are   --per_device_train_batch_size 4 \
                          --learning_rate 3e-5 \
                          --num_train_epochs 2 \
                          --max_seq_length 512 \
                          --doc_stride 250 \
                          --max_answer_length 200 \

## Evaluation
### Testing Data, Factors & Metrics
#### Testing Data
The model was trained and validated on train and validation sets.

#### Metrics
Here we use 2 metrics for QA tasks exact match and F-1.

### Results
{'exact_match': 37.12871287128713, 'f1': 64.90491019877854}

## Model Card Contact
Feel free to reach out to me at thelong20.4@gmail.com if you have any question or suggestion.