: Machine Learning - Research Paper Example

Early disease outbreaks detection will lead to improved public health (Saving of lives and reduction cost). An existing Bayesian non-temporal and non-spatial system is converted to a spatial temporal system. A system like PC (PANDA_CDCA) is converted to PANDA-CDCA-Temporal-Spatial (PCTS). It is then hypothesized that: (1) when an outbreak is up in a sub-region, PCTS will notice the occurrence earlier than PC; (2) once an outbreak is noticed, PCTS will maintain a detection indication that will PC without fail; (3) when an outbreak is up in a small sub-region, PCTS can calculate approximately that sub-region with more precision; and (4) PCTS will work as accurately as a SaTScan, which is a state-of-the-art spatial existing detection system.

Influenza and cryptosporidiosis outbreaks were injected and PC and PCTS were used comparatively to show how accurately they detected the injected disease (rather than detecting any just occurring outbreak). The results showed that the PCTS had greater sensitivity than PC at false alert rates. PCTS and SaTScan had 100% sensitivity for cryptosporidiosis at all false alert rates. Depending on false alert rate, SatScan-MT had a sensitivity of 5-10% improved than PCTS for Influenza detection. Method for Modeling of Unknown Diseases for Bio-surveillance This method uses Bayesian networks to model unknown causes of diseases (events) in the case of a disease outbreak.

The Bayesian approach detects and models both (1) known disease; using prior probabilities which are informative (2) Unknown diseases (e.g. a new contagious virus never seen before); using prior probabilities which are less informative (Shen and Cooper 589-593). An experiment was carried out using the BH algorithm in context of disease outbreak detection, Emergency Department (ED) records were also taken for use in the experiment. BH takes as input patient symptoms like cough, diarrhea and fever (most recent).

Models are then created i.e. Known Disease-specific Models (DSM); represents a person has a specific infection and Unknown-disease Models (UDM); represents a person has an unknown ailment. Three diseases (Cryptosporidiosis, inhalation tularemia and early stage anthrax) were chosen from CDC-A+ to be used in the experiment. Each of the 3 diseases was used to simulate an outbreak. In each of the experimental simulations (for each disease in the least), the three disease symptoms were modeled; abdominal pain, headache and cough.

The algorithm (HB) will take one symptom of the population as input at a time e.g. coughs vs. no cough. 2005 ED cases were obtained from a large hospital in Pennsylvia (Allegheny County). To determine false positive rates in the detection thresholds, the BH algorithm was run (using DSM1,UDM1, DSM2 and UDM2 models) on the ED cases (back ground time series) in 2005 which were assumed to not have disease outbreaks. For each model a threshold r was chosen that yielded one false-positive/month. Under the condition (one false positive per month) a threshold r was applied to the output probability ratios of outbreak scenario of a selected experiment resolve its detection time.

The mean detection time of all the disease models over the experiments was obtained as below (du denotes a distinct outbreak disease): DSM UDM Exp. 1 (du is modeled) 6.05 6.18 Exp.2 (du is not modeled) 7.14 6.38 Mean detection time (in days) of all