on 25-May-2022 (Wed)

The synthetic data that will populate the tables have to retain the properties of the original data. This is an additional challenge for the model since we have to preserve referential integrity, meaning that the foreign key — the column or group of columns that provide the link between two tables — in Table A has to match the corresponding items in Table B (if a relation is one-to-many). A possible solution to this problem is to synthesise data at multiple granularity levels:

1. Use unsupervised machine learning to cluster data at parent level (customer).

2. Synthesise this table, including the cluster identifier.

3. Randomly assign a synthesised customer to a real order sequence.

4. Finally synthesise the remaining variables (sequences at the child level) conditioned on the

previous data.

The problem with this solution is that it does not scale for very large databases.

Here we present an extension of the original GAN architecture to be more modular and hierarchical. In its simplest form, it consists of two modules, each composed of a generator and a discriminator, coupled in tandem, as illustrated in the figure below. Training of this architecture is done alternating:

Use Generator 1 to generate the common fields (customer name, location, etc) + Discriminator 1.
Use results from Generator 1 plus noise to conditionally generate the order details (product, quantity, etc) using a long short term memory network (LSTM) or convolutional neural network (CNN)

Advantage

The advantage of this method is that, contrary to the previous method, we don’t need to perform clustering and it is scalable to any number of hierarchical levels. Also, the approach is very flexible since we can have one or many parent tables and as many children as necessary.

hierarchical GAN

Data processing¶ The first step in data processing is to remove all personally identifiable information — like names and addresses — and any free-text fields.

We end up with the following attributes for the parent and child components:

| | | | | --- | --- | --- |Parent attributes | Name | Type | Description | | Employee_id | Categorical | 78 levels | | order_date | Datetime | 1996-1998 | | required_date | Datetime | 1996-1998 | | shipped_date | Datetime | 1996-1998 | | ship_via | Categorical | 3 levels | | freight | Numerical | [0.02 - 1007] | | ship_region | Categorical | 19 levels | | ship_country | Categorical | 21 levels | | customer_city | Categorical | 70 levels | | customer_region | Categorical | 19 levels | | customer_country | Categorical | 21 levels | | order_lenght | Numerical | [1 - 22] | | | | | | --- | --- | --- |Table 3: Child attributes Name Type Description product_id Categorical 77 levels supplier_id Categorical 29 levels category_id Categorical 8 levels unit_price Numerical [2 - 263.5] quantity_per_unit Numerical [0 - 70]

All data is converted into a one-hot encoding, including datetime and numerical types (previously binned into 20 levels). One-hot encoding converts categorical variables (occupation, city names, etc) into a numeric format to be ingested by machine learning algorithms.

The final generated tabular data is converted back into tables to recreate the original database

hierarchical GAN

Training

Machine learning training is performed using an optimised set of hyperparameters for the GAN architecture. Instead of alternating the training between Module 1 and 2, we opted to first train Module 1 (GAN1) and then Module 2 (GAN2). The loss function used was the Wasserstein loss.

hierarchical GAN - Summary

There are many ways of dealing with this problem using different levels of granularity. We presented one that can be easily extended to handle temporal sequences. Its main advantages are modularity and scalability as it can be extended to multiple granularity levels.

The main drawback of this approach is that it may become difficult to implement if the database contains too many one to many relations and nested hierarchical levels.

Traditional methods of synthetic data generation use techniques that do not intend to replicate important statistical properties of the original data. Properties such as the distribution, the patterns or the correlation between variables, are often omitted. Moreover, most of the existing tools and approaches require a great deal of user-defined rules and do not make use of advanced techniques like Machine Learning or Deep Learning. While Machine Learning is an innovative area of Artificial Intelligence and Computer Science that uses statistical techniques to give computers the ability to learn from data, Deep Learning is a closely related field based on learning data representations, which may serve useful for the task of synthetic data generation

The clinical presentation depends on the child’s age. The younger the child, the fewer and more atypical the clinical signs and symptoms [1]. Questions asked to parents aim to identify the recent onset of signs and symptoms. Some may be observed very early on: digestive signs are common and often misleading (vomiting, diarrhea); one third of infants present with behavioral disorders (recent change in behavior, fearful child or inconsolable crying, lower consciousness, difficulties in bottle feeding); focal or gen- eralized seizures are observed in 20–30% of infants, sometimes with status epilepticus [2]. Fever may not be observed in young infants [3].

Fever was the first symptom in children aged below 5 years. Irrespective of age, the first truly specific clinical signs were signs of sepsis: abnormal skin coloration, cold extremities, muscle pain in the legs

The first clinical sign indicative of meningococcal infection was often a non-specific skin rash. Within a few hours, petechiae were observed followed by an extensive hemorrhagic rash approximately 12 hours after symptom onset. Median time to onset of early meningitis typical signs (neck stiffness, photophobia, and bulging fontanelle — anterior when still open — [usually before 6 months of age]) was 12–15 hours. Median time to onset of late signs (consciousness disorders, seizures) was 15 hours in infants aged below one year and approximately 24 hours in older infants

The authors of a prospec- tive study [2] of 3066 infants aged below 3 months presenting with fever and managed at a community pediatrician office in the United States, identified the most specific signs of severe sepsis and pointed out three factors: the child’s general presentation (change in skin color, deterioration of general status), behav- ioral changes (responsiveness and interactivity disorders, loss of smile), and fever > 39.5 ◦ C. The physician’s clinical judgment was in this study the most decisive factor. Hospital surveillance is recommended when one of these signs is observed in infants aged below 3 months.

A lumbar puncture should be performed in infants aged below 3 months if one of the following signs is observed: • unusual behavior (plaintive crying, moaning, inconsolable crying, hyporeactivity, irritability, pain, cutaneous hyperes- thesia); • tachycardia with normal blood pressure, time to skin recol- oration > 3 seconds, cyanosis; • neurological abnormalities (bulging fontanelle, neck hypoto- nia, general hypotonia) — Kernig’s and Brudzinski’s signs are usually not observed; • seizures; • purpura

Typical clinical symptoms (vomiting, bulging fontanelle, neck stiffness, photophobia, consciousness disorders) are most common in infants aged between 3 months and 2 years, although they can be absent. A lumbar puncture is required when seizures are observed in feverish children aged below 6 months and should be discussed in feverish children aged between 6 and 12 months

Older children present with signs and symptoms more similar to those observed in adults: fever, chills, vomiting, photophobia, severe cephalgia. Several signs are associated with more specific etiologies: purpura and joint disorders are indicative of Neisseria meningitidis infection; otorrhea or otitis and a history of brain injury are indicative of a breach in the meninges and of pneumo- coccal infection. Signs and symptoms may first be observed on an occasional basis with recurrent seizures in patients presenting with fever, consciousness disorder, coma, and status epilepticus

The authors of a meta-analysis of articles published between 1966 and 1997 assessing clinical signs of meningitis and includ- ing 733 patients (90% of bacterial meningitis cases), reported 46% sensitivity for the triad “fever, neck stiffness, and altered state of consciousness (or cephalgia)”. However, 95% of patients presented with at least two of these symptoms and 100% with at least one. The meningitis diagnosis could therefore be ruled out in patients who did not present with any of these three signs [7].

Recent cohort studies of bacterial meningitis confirmed that fever was observed in 77–97% of cases, cephalgia in 58–87% of cases, neck stiffness in 65–83% of cases, and consciousness disorders in 30–69% of cases. The triad of fever, consciousness disorder, and neck stiffness was observed in 41–51% of cases [8–12]

A Japanese study suggested a new clinical sign: jolt accentuation. The aim is to identify exacerbation of cephalgia with rapid rotations of the head in the horizontal axis (2 to 3 per second) [15]. Sensitiv- ity of jolt accentuation is 97%, its specificity 60%, its positive predictive value 81%, and its negative predictive value 92%.

Several studies reported the poor sensitivity of Kernig’s and Brudzinski’s signs and of neck stiffness

The largest prospective study conducted with adults [12] included 696 episodes of community-acquired bacterial meningitis (51% due to S. pneumoniae and 37% due to N. menin- gitidis). The triad of “fever, neck stiffness, and altered state of consciousness” was associated with 44% sensitivity. However, 95% of patients presented with at least two signs among cephal- gia (87%), fever (77%), neck stiffness (83%), and altered state of consciousness (Glasgow score < 14 in 69% of patients, includ- ing coma in 14%). Focal neurological signs were observed in 33% of patients and seizures in 5% of patients

The usual triad of symptoms was more commonly observed in patients present- ing with pneumococcal meningitis than in those presenting with meningococcal meningitis (58% versus 27%, P < 0.001)

Cuta- neous signs, mainly purpura, were observed in 26% of episodes (meningococcal meningitis: 95% of cases; otherwise pneumo- coccal meningitis)

meningitis is highly probable when patients present with fever and purpura, even more so when cephalgia is also observed

particular attention should be paid to patients with risk factors for meningitis such as those with chronic excessive alco- hol consumption, psychiatric disorders, and homeless people. The meningitis diagnosis can be extremely difficult to estab- lish in these patients on the sole basis of clinical presentation

The lumbar puncture should be performed within one hour following admission to the emergency department, unless contraindicated or provided an imaging test is required before- hand

When the lumbar puncture cannot be immediately performed, an empirical antibiotic therapy should be initiated, dexametha- sone should be administered, and at least two blood cultures should be performed beforehand (Fig. 1).

non-neurological contraindications: ◦ cutaneous infection localized to the lumbar puncture site, ◦ uncontrolled hemodynamic or respiratory disorder, ◦ known hemostasis disorders (hemophilia, other coagu- lopathy, platelet count < 50,000/mm 3 ); OR anticoagulant treatment (effective doses of vitamin K antagonists; unfractionated or fractionated heparin or direct oral anti- coagulants, irrespective of the dose); OR spontaneous bleedings indicative of disseminated intravascular coagu- lation (DIC) [16] (Table 1).

An antiplatelet treatment does not contraindicate the lumbar puncture [17,18].

The CSF should be collected into four sterile tubes for bio- chemical, microbiological, and cytological analysis purposes and for bacterial identification using molecular biology tech- niques. The first tube should not be used for these analyses to prevent any risk of contamination.

The overall quantity of CSF required is 2 to 5 ml in adults (40–100 drops), ideally 2 ml (approximately 40 drops) in children, and approximately 0.5 ml (10 drops) for the molecular biology analysis

Bacterial detec- tion and isolation depend on the volume used for analysis, but also on compliance with transport conditions and time require- ments. The tube for molecular biology should be kept in frozen conditions (−20 ◦ C or −80 ◦ C) if transport to the laboratory is delayed.

Results of the cytological and biochemical anal- yses, Gram staining and, if relevant, antigen detection by immunochromatographic assay should be communicated to the patient’s medical team within the hour following lumbar puncture