Pré-Publication, Document De Travail Année : 2024

Unveiling Mosquito Patterns in Chicago (2007-2024): A Data Analytics and Machine Learning Study

Ilyas Dr Potamitis
  • Fonction : Auteur
  • PersonId : 1435189

Résumé

We apply data analytics to the publicly available and recently updated Chicago 2007-2024 Mosquito Database. In this database, 195 traps have been deployed in Chicago, Illinois, USA, from 2007 to 2024. Every year, from late May to early October, public health workers in Chicago set up mosquito traps scattered across the city. These traps collect mosquitoes, which are then partitioned into batches of fifty specimens. Each batch has been assessed using Polymerase Chain Reaction (PCR) for the presence of West Nile virus before the end of each week. The database records include the number of mosquitoes, the mosquito species, geographical information, and whether West Nile virus is present in each cohort. In its first part, this work explores the application of mosquito data analytics to the manually collected data, focusing on the potential to identify trends, find the outbreaks, and localize hotspots to support vector control strategies. In its second part, we investigate at what extent a virus-positive batch can be predicted using the rest of the variables recorded in the database, showing that an AUC score of approximately 81% can be achieved on a 2-year held out subset without including weather data. Finally, we discuss our findings in the context of integrating automated insect counting traps (etraps) with mosquito data analytics. We argue that optical counters with environmental sensors embedded in traps can provide the supplementary information used in the Chicago database to predict the probability of an infected cohort without the use of PCR analysis. The probability of an infested cohort is less accurate than PCR but comes at no extra cost and is delivered almost real-time contributing to public awareness and resource allocation to intervention activities.

Recent advances in machine learning and spatial analysis have further enhanced the capabilities of mosquito data analytics (see [9]-[10] and [11] for statistical challenges). Spatial analysis tools such as geographic information systems (GIS) have enabled researchers to map high-risk areas and understand how environmental factors contribute to mosquito breeding and disease spread [12-13]. The use of statistics is particularly important to understand mosquito surveillance Data in Arizona [14] and elsewhere, where recent studies have utilized detailed mosquito trapping and WNV occurrence data to identify outbreaks and guide targeted interventions [15]. Data analytics and machine learning has been applied on mosquito-related dataset in various contexts [16-19]. Research in [20] models the distribution of invasive mosquito species using several machine learning techniques on tabular datasets. In [21], the authors use climate data to predict malaria incidence, which is linked to mosquito populations. In [22], researchers explore the use of tabular data to forecast mosquito vector abundance. In [23], machine learning models are applied to mosquito occurrence data, analyzing mosquito habitat based on regional climate data. In [24-25], data mining and machine learning techniques are used to understand relationships among vectors, hosts, and pathogens. The established procedure for identifying mosquitoes with a virus load is to subject them to PCR testing. PCR is a molecular biology technique used to detect the presence of specific pathogens or viruses in mosquito samples. It works by amplifying small segments of DNA or RNA, allowing researchers to identify and confirm the presence of disease-causing agents, such as West Nile Virus, Dengue Virus, or Malaria Plasmodium in mosquitoes. While PCR is highly effective for detecting pathogens, it has several practical disadvantages: (a) PCR requires specialized reagents (such as enzymes, primers, and nucleotides) and consumables (e.g., tubes and plates). The cost per test can add up significantly, making it expensive for large-scale mosquito surveillance programs. (b) PCR requires skilled personnel and sophisticated laboratory equipment, such as thermal cyclers, which are costly to purchase and maintain, especially in under-resourced regions. (c) PCR is not a real-time monitoring tool; the process involves collection, transportation to a lab, sample preparation, and testing, which introduces delays. It may take days or weeks to process and analyze samples from the field, leading to a lag between data collection and actionable results. Near-infrared (NIR) spectrometry has been suggested as an alternative approach for virus detection in mosquitoes. Although it relaxes some of the strict requirements of PCR, such as reagent use, it still requires specialized personnel and costly equipment [26-30]. NIR spectrometry is faster than PCR but not instantaneous, and it requires careful placement of the sensing probe on a mosquito specimen, making it unsuitable for automated analysis of large numbers of mosquitoes. Although is a strong statement, we argue that traditional mosquito surveillance practices are time-consuming, expensive, and lack scalability [31]. In this work, we are mainly interested in investigating whether we can predict the probability of an infected WNV batch in mosquito traps based on other variables such as the date, location, number of batches per trap, and number of mosquitoes per batch given historical data with manually verified virus presence. Machine learning models have been employed to predict mosquito populations and disease outbreaks with high accuracy, often outperforming traditional statistical approaches. This work seeks to explore the growing role of mosquito data analytics and machine learning on the publicly available, tabular dataset of Chicago Mosquito records (2007-2024) in addressing public health challenges posed by mosquito-borne diseases. While our findings indicate that the probability attributed to each batch of being infected is not as accurate as PCR, it is a cost-effective and instantaneous approach. We then discuss the technical challenges that must be overcome so that automated optical counters embedded in mosquito traps [32-38], which can extract the variables used in this study, could be adapted to report an informed probability of a WNV-positive cohort. Additionally, in terms of resource allocation, which is always an issue in practice, we suggest that it is more effective to allocate PCR analysis to the cohorts flagged as positive by automated traps. We open-source the code used to analyze the public data and classify the Chicago mosquito (2007-2024) database, making it applicable to any mosquito database with a similar structure (see Appendix).

The database's fields

The database is tabular, and it is important to note that each row corresponds to a batch of mosquitoes and several rows can belong to the same trap visit as the catches are partitioned in groups of fifty specimens. The database is highly unbalanced as less than 10% of the batches have a WNV positive label. The mosquito occurrences dataset contains the following columns: SEASON YEAR: The year of data collection. WEEK: The week of the year that has been assessed with PCR. TEST ID: Unique identifier. BLOCK: General location of the mosquito trap

Fichier principal
Vignette du fichier
Chicago_Mosquitos_Data_analytics.pdf (2.04 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-04763207 , version 1 (01-11-2024)

Licence

Identifiants

  • HAL Id : hal-04763207 , version 1

Citer

Ilyas Dr Potamitis. Unveiling Mosquito Patterns in Chicago (2007-2024): A Data Analytics and Machine Learning Study. 2024. ⟨hal-04763207⟩
140 Consultations
208 Téléchargements

Partager

  • More