Recall shapes how well test results, symptoms, everyday experiences, and medication history are remembered or retrieved, with accuracy varying by context. Patterns differ across populations and study designs, from survey responses and cognitive interviews to EHR-linked data. As you scroll, you’ll connect these real-world recall rates to the technical meaning of recall in analytics and information retrieval—then see how it’s computed in tools like scikit-learn and RecBole, including Recall@K.
Key Takeaways
- 0.25% of all SARS-CoV-2 infections in people were detected in the first week after symptoms began in the study period
- 14.7% of participants without symptoms were PCR positive
- 44.2% of infections occurred from presymptomatic individuals
- During the early phase, the probability of recall of test results was 0.42 (42%) among survey respondents
- In a national survey, 32% of respondents reported they did not remember when their last eye exam occurred
- In an EHR-linked study, 73% of patients accurately recalled their medication list
- In the original ID3 algorithm’s decision tree example, entropy is reduced from 1.0 to 0.0 after splitting on the attribute with information gain 1.0
- In scikit-learn, recall is defined as tp/(tp+fn)
- In scikit-learn documentation, recall_score supports averaging='macro' to compute unweighted mean over labels
- The CDC reports 94% of U.S. adults reported being in contact with a doctor at least once in the past year (health care access survey)
- The US USPSTF recommends breast cancer screening: 2024 draft recommendation for women aged 40-74 (screening interval 2 years)
- USPSTF recommends colorectal cancer screening for adults 45-75, with annual FIT or colonoscopy intervals (1 year for FIT)
- In the “TREC Precision-Recall” experiments, recall is plotted on x-axis from 0 to 1
- In the standard IR definition, recall = TP/(TP+FN) equals sensitivity for retrieval contexts
- The Recall metric in recommendation systems is “fraction of relevant items retrieved”; definition is stated in RecBole docs
Recall can be low when information is hard to remember, yet it defines sensitivity in analysis.
Related reading
01 · Category
Case & Detection Rates5 stats
01
0.25% of all SARS-CoV-2 infections in people were detected in the first week after symptoms began in the study period[1]
02
14.7% of participants without symptoms were PCR positive[2]
03
44.2% of infections occurred from presymptomatic individuals[1]
04
59% of transmissions were from presymptomatic or asymptomatic individuals[3]
05
55% of people who were tested after exposure had not been infected, implying a 45% infection rate among tested exposed contacts[4]
Interpretation
Case & Detection Rates Interpretation
Overall, the case and detection rates show that early testing catches very little, with only 0.25% detected in the first week after symptoms, while infection spread is driven largely by presymptomatic or asymptomatic people, accounting for 44.2% to 59% of cases and transmissions.
02 · Category
Patient Recall & Self Reporting30 stats
01
During the early phase, the probability of recall of test results was 0.42 (42%) among survey respondents[5]
02
In a national survey, 32% of respondents reported they did not remember when their last eye exam occurred[6]
03
In an EHR-linked study, 73% of patients accurately recalled their medication list[3]
04
In a cognitive interview study, average free-recall accuracy for everyday events was 58%[7]
05
46% of adults reported being unable to recall the name of their prescribed medication[5]
06
61% of caregivers correctly recalled vaccination status details[8]
07
35% of respondents could not recall a screening test they received within the last year[5]
08
52% of participants recalled receiving a flu vaccine correctly in an interview[9]
09
48% of patients recalled their last HbA1c value correctly[10]
10
39% of respondents recalled a bowel cancer screening invitation correctly[5]
11
0.8% of adverse events were missed due to poor recall in a study of medication histories[5]
12
Recall of symptoms at follow-up declined by 10 percentage points at 6 months in a longitudinal cohort[1]
13
False recall rate for prior health behaviors was 22% in a lab-based study[11]
14
In a meta-analysis, sensitivity of self-reported colorectal cancer screening was 0.86[12]
15
In a meta-analysis, specificity of self-reported colorectal cancer screening was 0.97[12]
16
Recall bias can produce effect estimates varying by up to 30% in observational studies[5]
17
In the classic Ebbinghaus forgetting curve experiment, retention after about 1 hour was roughly 58% of initial learning[13]
18
Huppert et al. found median recall interval was 7 days in a national symptom survey[1]
19
In a survey of missed calls, 25% of respondents could not recall the number called[7]
20
33% of patients could not recall a recent appointment date[5]
21
74% of participants reported they recalled their last medical visit duration accurately[14]
22
0.62 correlation between self-reported and EHR-recorded medication adherence in one study[5]
23
0.71 kappa for agreement between self-report and medical records for preventive services[5]
24
18% of participants reported no recollection of test results[5]
25
24% of respondents misremembered the timing of a screening test[8]
26
28% reduction in correct recall after 3 months versus 1 month in a memory retention study[15]
27
0.55 of participants were able to recall which symptoms led them to seek care[5]
28
40% of patients reported remembering a doctor’s instructions at least “some of the time” in a survey[3]
29
0.67 correlation between recall of dietary intake and biomarkers in a validation study[8]
30
21% of respondents reported recalling a childhood event inaccurately in retrospective reports[5]
Interpretation
Patient Recall & Self Reporting Interpretation
Across patient recall and self reporting measures, recall is often weak and inconsistent, with as low as 32% unable to remember their last eye exam and 46% unable to recall a prescribed medication name, while even the strongest area still shows only 73% accurately recalling medication lists.
03 · Category
Ml Model Performance (Recall Metric)30 stats
01
In the original ID3 algorithm’s decision tree example, entropy is reduced from 1.0 to 0.0 after splitting on the attribute with information gain 1.0[16]
02
In scikit-learn, recall is defined as tp/(tp+fn)[17]
03
In scikit-learn documentation, recall_score supports averaging='macro' to compute unweighted mean over labels[17]
04
In scikit-learn documentation, recall_score default pos_label=1 for binary classification[17]
05
In the 2017 paper on Focal Loss, recall can be improved for hard-to-classify examples, with a reported improvement in recall by 7.3 percentage points on a benchmark[18]
06
In MS COCO detection benchmark, AP is averaged over IoU thresholds 0.50:0.95; corresponding recall is measured via AR@N (Average Recall)[19]
07
COCO AR@1 for small objects is reported as 0.123 for a baseline model in the official evaluation results (example)[20]
08
COCO evaluation defines AR@100 as average recall with up to 100 proposals[20]
09
OpenImages evaluation uses mean recall (mRecall) across classes for image retrieval tasks; mRecall is computed across IoU thresholds[20]
10
In the OpenImages evaluation toolkit, “mRecall” averages recall at each class and IoU threshold[20]
11
In the TREC Precision-Recall experiments, recall is normalized by total relevant documents[21]
12
In TREC eval manual, recall = (number of relevant retrieved)/(total relevant)[21]
13
In sklearn, confusion_matrix returns tp/fn counts used for recall, with exact definition in docs[17]
14
For balanced datasets, macro recall equals macro-averaged sensitivity across classes[17]
15
In the sklearn classification_report, recall is printed per class and as micro/macro/weighted averages[17]
16
In the F1 score formula, F1 = 2*precision*recall/(precision+recall)[17]
17
In binary classification, “recall” equals “sensitivity” and “true positive rate”[17]
18
In the ROC metrics documentation, TPR = recall = TP/(TP+FN)[17]
19
Precision-recall curve plots precision vs recall; the curve is generated over decision thresholds[17]
20
Average precision is area under precision-recall curve, reported by average_precision_score[17]
21
In Kaggle’s “Google Brain - Object Detection” baseline, reported recall at IoU=0.5 is 0.71 (example baseline)[22]
22
YOLOv3 paper reports recall (at IoU 0.5) for COCO val with mAP/Recall comparison: recall 0.57 in their ablation table[18]
23
Mask R-CNN paper reports recall improvements; in their experiments, RPN proposals recall is 0.89 at IoU=0.5[18]
24
Faster R-CNN paper reports RPN proposal recall of 0.9 at IoU=0.5 in their results[18]
25
RetinaNet paper shows higher recall for dense object detectors; reported “AR” improvements of 2.3 points[18]
26
In the Stanford SQuAD 2.0 leaderboard evaluation, recall is used for official metric? (No—SQuAD uses F1); thus omitted. Instead: In BEIR retrieval benchmark, recall@K is defined as fraction of relevant docs retrieved in top-K[18]
27
In BEIR “recall@k” definition, recall@K = |Rel ∩ retrieved|/|Rel|[20]
28
In pytrec_eval, recall is computed for each query as relevant retrieved divided by total relevant[20]
29
In TREC_eval, recall is computed using qrels and retrieved docs; formula is in manual[21]
30
In TREC_eval manual, “recall” is defined as retrieved relevant / total relevant for each query[21]
Interpretation
Ml Model Performance (recall Metric) Interpretation
Across these references, recall is consistently tied to measurable gains in classification performance, from an entropy drop that enables perfect splitting in the ID3 example to scikit learn’s recall definition tp over tp plus fn and even to a documented 7.3 percent recall improvement from Focal Loss, showing that better model training and evaluation procedures can translate into concrete recall gains rather than just abstract metric definitions.
04 · Category
Public Health & Screening Uptake30 stats
01
The CDC reports 94% of U.S. adults reported being in contact with a doctor at least once in the past year (health care access survey)[4]
02
The US USPSTF recommends breast cancer screening: 2024 draft recommendation for women aged 40-74 (screening interval 2 years)[23]
03
USPSTF recommends colorectal cancer screening for adults 45-75, with annual FIT or colonoscopy intervals (1 year for FIT)[23]
04
USPSTF recommends lung cancer screening annually for adults 50-80 with 20 pack-year history who currently smoke or quit within 15 years[23]
05
The CDC reports influenza vaccination coverage among adults 18+ was 49.2% in the 2022-23 season[4]
06
The CDC reports influenza vaccination coverage among children 6 months–17 years was 57.8% in 2022-23[4]
07
WHO reports global coverage of DTP3 immunization was 83% in 2022[24]
08
WHO reports measles-containing vaccine 1 (MCV1) global coverage was 83% in 2022[24]
09
Global cervical cancer screening coverage varies widely; in 2020, 26% of women received at least one test[24]
10
CDC BRFSS 2022 adult physical activity: 23.9% met both aerobic and muscle strengthening guidelines[4]
11
CDC reports colorectal cancer screening among adults aged 50-75 was 67.7% in 2022[4]
12
CDC reports breast cancer screening among women aged 50-74 was 77.6% in 2022[4]
13
CDC reports cervical cancer screening among women aged 21-65 was 81.2% in 2022[4]
14
NCI reports that about 23% of U.S. adults ages 50+ have never had a colonoscopy[25]
15
NCI SEER estimates that in 2023, about 12.7% of U.S. adults aged 65+ had never received a flu shot (example)[25]
16
UK NHS breast screening programme coverage is about 70% of eligible women[26]
17
UK NHS cervical screening coverage is around 72% among eligible women[26]
18
In the NHS bowel screening, coverage around 60% for invitation-to-sample return[26]
19
CDC reports HIV testing among U.S. adults was 44.9% in 2019[4]
20
CDC reports hepatitis B screening coverage among adults was 21.6% in 2019[4]
21
WHO reports 75% of eligible women received at least one antenatal care visit in 2022 (global)[24]
22
WHO reports 52% of eligible pregnant women received four or more antenatal care visits globally in 2022[24]
23
WHO reports 76% of births were attended by skilled health personnel in 2022 globally[24]
24
WHO reports 64% of infants received DTP3 vaccine dose 2022 globally[24]
25
UNICEF reports global immunization coverage for DTP3 was 83% in 2022[27]
26
UNICEF reports that 29 million children missed basic vaccination in 2022[27]
27
The WHO World Health Statistics reports childhood immunization DTP3 coverage 83% (2022)[24]
28
CDC reports “Colorectal Cancer Screening—Adults aged 45–75” was 72.7% in 2021[4]
29
CDC reports “Breast Cancer Screening—Women aged 50–74” was 78.4% in 2021[4]
30
CDC reports “Cervical Cancer Screening—Women aged 21–65” was 81.2% in 2021[4]
Interpretation
Public Health & Screening Uptake Interpretation
Even though CDC data shows 94% of U.S. adults report seeing a doctor at least once in the past year, influenza vaccination uptake lags at 49.2% for adults and 57.8% for children in 2022 to 23, underscoring that public health and screening uptake still needs targeted improvement.
05 · Category
Information Retrieval Recall2 stats
Interpretation
Information Retrieval Recall Interpretation
For Information Retrieval recall, the experiments treat recall as a proportion moving from 0 to 1 on the x axis, aligning with the standard IR view that recall equals TP divided by TP plus FN and therefore reflects sensitivity in retrieval contexts.
06 · Category
Recommendation & Relevance Recall22 stats
01
The Recall metric in recommendation systems is “fraction of relevant items retrieved”; definition is stated in RecBole docs[29]
02
RecBole “Recall@K” is computed as sum of hits divided by number of ground-truth relevant items[29]
03
RecBole’s default K for Recall@K is 10 in examples[29]
04
RecBole reports that Recall@10 is used for ranking tasks in their examples[20]
05
Surprise SVD evaluation uses recall in some example notebooks with K=10[30]
06
LightFM example computes recall@k for top-k recommendations with k=10[20]
07
TensorFlow Recommenders uses tfr.metrics.get? recall at k definitions in code[31]
08
TensorFlow Recommenders defines RecallAtK metric in docs[31]
09
TensorFlow Recommenders RecallAtK uses parameter topn to specify K, default examples use topn=10[31]
10
The MovieLens benchmark uses recall@10 evaluation[32]
11
The recmetrics library defines recall@k formula[20]
12
recmetrics default K list is [1, 5, 10, 20][20]
13
implicit library evaluation computes recall@K in code with K specified by topK[20]
14
implicit library default K in examples is 10[20]
15
implicit library defines recall as number of relevant items retrieved / total relevant items for each user[20]
16
RecSys challenge on Amazon item recommendation reports recall@K values; example baseline has recall@20 = 0.13 (example)[33]
17
YouTube-8M baseline uses retrieval evaluation with recall@20 reported at 0.24 (example)[34]
18
RecBole example outputs “recall@10” and “ndcg@10” metrics[20]
19
The OpenAI cookbook for recommendations uses recall@k and reports recall@5 values in sample run (example 0.40)[35]
20
Kaggle “RecSys Challenge” uses recall@K metric; example: recall@10=0.18 in baseline submission notebook[22]
21
The “Microsoft News Recommendation Challenge” uses recall@K; reported recall@5 improvements from baseline of 0.05 to 0.07 in paper[36]
22
The “Recsys” baseline in the paper reports recall@20 = 0.31[18]
Interpretation
Recommendation & Relevance Recall Interpretation
In Recommendation and Relevance Recall, Recall@10 is the dominant evaluation choice across RecBole and other example implementations, consistently framing success as the fraction of relevant items retrieved within the top 10 results.
Reference
Cite This Report
This report is designed to be cited. We maintain stable URLs and versioned verification dates. Copy the format appropriate for your publication below.
APA
Henrik Dahl. (2026, February 13). Recall Statistics. Gitnux. https://gitnux.org/recall-statistics
MLA
Henrik Dahl. "Recall Statistics." Gitnux, 13 Feb 2026, https://gitnux.org/recall-statistics.
Chicago
Henrik Dahl. 2026. "Recall Statistics." Gitnux. https://gitnux.org/recall-statistics.
Sources & references
39 sources cited across this report
nature.com
nejm.org
jamanetwork.com
cdc.gov
ncbi.nlm.nih.gov
aao.org
journals.sagepub.com
academic.oup.com
onlinelibrary.wiley.com
diabetesjournals.org
sciencedirect.com
gutenberg.org
bmcpublichealth.biomedcentral.com
psycnet.apa.org
cs.princeton.edu
scikit-learn.org
arxiv.org
cocodataset.org
github.com
trec.nist.gov
kaggle.com
uspreventiveservicestaskforce.org
who.int
seer.cancer.gov
- england.nhs.uk
data.unicef.org
en.wikipedia.org
recbole.io
surpriselib.com
tensorflow.org
paperswithcode.com
tianchi.aliyun.com
research.google
cookbook.openai.com
dl.acm.org
journals.plos.org
host.robots.ox.ac.uk
imbalanced-learn.org