Early Detection of Dengue Infection Using Machine Learning: A Survey Dataset from Bangladesh

Description

This dataset contains 1,018 questionnaire responses collected in Bangladesh for a research study on early detection of dengue infection using machine learning. Responses were recorded through a bilingual English–Bengali Google Forms questionnaire. The researcher has confirmed this as the final collection. The survey covers age, gender, district and residence type; reported dengue history and environmental or household exposure; fever duration and maximum recorded temperature; headache, retro-orbital, muscle and joint pain rated on 0–10 scales; vomiting frequency; eight symptoms reported during illness days 1–3; and the recorded final dengue test result. There are 630 recorded positive and 388 recorded negative outcomes. The recorded test outcomes have not been independently verified against laboratory or clinical records. Test modality and illness day of testing are not specified in the supplied data. Dengue_Raw_Research_Data_v1.csv contains 1,018 rows and 23 research columns, preserving the original research values, missing entries, anomalies, outcome labels and row order. Three administrative fields—submission timestamp, participation-consent response and data-entry role—are excluded from both public CSVs. Short English headings are mapped to the original bilingual questions in Data_Dictionary.csv. Dengue_Cleaned_Research_Data_v1.csv contains the same responses and all 23 research fields, including every symptom and pain-scale variable. Cleaning standardizes headings, category labels, district spelling and surrounding whitespace. The cleaned file has zero missing cells: 12 source temperature blanks and two invalid entries were filled with 103°F, the median of 1,004 valid recorded temperatures. These 14 estimates are individually traceable in Cleaning_Log.csv; original values remain in raw. All recorded valid measurements, pain ratings, outcomes and row order are preserved. Two recorded values of 107°F are retained and flagged. No rows were removed or generated. No scaling or model-specific feature selection was applied. Data_Dictionary.csv, Cleaning_Log.csv and README.md document definitions, transformations, imputation positions, uncertain responses, quality checks and hashes. README.md includes a tested reconstruction example. For machine-learning evaluation, restore the 14 imputed positions to missing before splitting and fit imputation within training folds. Assess predictor availability at the intended prediction time, since some questions refer to the whole illness rather than its first three days.

Steps to reproduce

1. Read README.md and Data_Dictionary.csv. The original survey export has 1,018 responses and 26 columns. The public Dengue_Raw_Research_Data_v1.csv preserves the 23 research fields and their row order. The original 26-column export remains unchanged in the researcher's source records. 2. Exclude only the timestamp, participation-consent response and data-entry role from both public CSVs. Retain all 23 research fields in their original order, using the short headings in the dictionary. Keep all records and preserve the original outcome strings. 3. Apply Unicode NFKC normalization and trim surrounding whitespace in working predictor values. Standardize category labels with explicit mappings. Keep uncertain exposure responses as unknown rather than no. Normalize district case and whitespace and apply the explicit spelling mappings in Cleaning_Log.csv. 4. Parse numeric fields using the stated bounds. Preserve valid values and all four 0–10 pain ratings, including zero. Identify 12 source temperature blanks and treat the two invalid entries, 8 and Munshiganj, as missing in working data. Calculate the median from 1,004 valid recorded temperatures: 103°F. Fill all 14 working temperature gaps with that median and record each position in Cleaning_Log.csv. Keep raw unchanged and retain the two recorded 107°F entries with review flags. Do not deduplicate or scale the data. 5. To reconstruct Dengue_Cleaned_Research_Data_v1.csv exactly, copy the public raw data and apply only Cleaning_Log.csv entries with entry_type=value_change at the listed one-based response_positions, checking each raw_value before assigning cleaned_value. Skip informational entries. The Python standard-library example in README.md reproduces the cleaned file in memory and verifies every cell and file byte without overwriting the data. 6. Verify dimensions of 1,018 × 23 for both CSVs; 630 positive and 388 negative outcomes in both; 12 original missing raw temperature cells, both original invalid raw entries, and zero missing cleaned cells; exactly 14 median-imputed temperatures; zero exact duplicate rows in either file; and unchanged row order. Cleaning_Log.csv accounts for every altered retained cell. Use the SHA-256 values in README.md to verify the four CSV files.

Institutions

• Daffodil International University

  Dhaka Division, Dhaka

Categories

Licence