Accuracy of plant identification applications to identify plants in suspected poisoning cases referred to the Queensland Poisons Information Centre

Abstract

Accidental exposure to potentially poisonous plants is a common reason for contacting poisons information centres. Although most exposures are mild, early identification can be critical for severe cases. Smartphone apps that use image recognition can serve as an alternative to traditional botanist identification. This study investigates the accuracy of smartphone applications compared to botanist identification in real-life plant exposure cases referred to the Queensland Poisons Information Centre. This prospective observational study analysed data from 25 plant exposure cases between February 2022 and January 2023. In these cases, images were referred to a botanist for identification and reviewed by six researchers using six different plant identification applications (Google Lens™, PictureThis, PlantSnap, LeafSnap, Plant Identifier, and Pl@ntnet™). The researchers’ identifications were then compared to the botanist’s identification. The accuracy of the applications varied considerably, PictureThis was the best, with 74% correct identification per exposure, followed by Plant Identifier and Pl@ntnet™ (both 72%). PlantSnap had the lowest accuracy (38%). Agreement between researchers using the same application also showed variation. Current smartphone plant identification applications are not reliable enough for poison information centres to use in guiding advice for plant exposure calls. Botanist identification remains the gold standard. Clinicians should be wary of their use in real-world scenarios.

Keywords:

• Toxicology
• smartphone applications
• plant identification
• poisons information

Introduction

Contact with potentially poisonous plants is a frequent reason for calling a Poisons Information Centre [1,2]. While most exposures cause minimal or mild symptoms, early identification may help identify those plants associated with more severe toxicity [3–5]. Accurate identification of the plants involved can help guide poison information specialists in deciding which cases should be attended for medical assessment and who can safely stay at home [6]. In some instances, the plant species is known or identifiable. When uncertainty exists and images are available, these cases may be referred to local botanists for identification. However, formal botanist identification is not available at all times.

Technological advances have been made in plant identification using computer vision and deep learning algorithms. This brings the opportunity to use smartphone applications to aid with species recognition [7–9]. These applications are readily available and could provide an alternative when a botanist is unavailable. However, their accuracy is not established. Two prior studies of photographs of plants taken for research purposes found that the applications had variable accuracy between 53% and 96% but misidentified several toxic species [10,11]. Another study that looked at the accuracy of PictureThis in identifying 369 images of poisonous plants taken from a textbook found the accuracy to be 82% in identifying the correct genus [12]. Neither study used images of plants involved in real-life poisoning exposures.

We aimed to investigate the accuracy of smartphone plant identification applications compared to the gold standard of botanist identification in plant exposures referred to the Queensland Poisons Information Centre (QPIC). The results of this study will help determine whether smartphone plant identification applications can identify plant species with sufficient accuracy to provide useful clinical advice when specialised botanist advice is unavailable.

Method

Study design and setting

This was a prospective observational series, from February 2022 to January 2023, of plant-related calls to the Queensland Poisons Information Centre that were referred to a botanist for identification. The Queensland Poisons Information Centre receives approximately 900 plant exposure-related calls annually, which represents 3% of all calls [6]. Approximately 60 of these calls are referred to the Queensland Herbarium for expert identification by a botanist [6].

Selection of participants and intervention

Any plant-related exposures, when a plant image or images were referred to the Queensland Herbarium and an identification made were included in the study. Data collected included patient demographics (age, sex), exposure details (route of exposure, intent, Poisoning Severity Score (PSS) [13], poisons advice given, image of plant implicated and the botanist identification. Six toxicology clinicians (three clinical toxicologists, two clinical toxicology fellows and a clinical toxicology nurse practitioner), blinded to the botanist’s identification, were asked to review the images using a plant identification application to determine the plant’s identity. Six separate plant identification applications were studied: Google Lens™, PictureThis, PlantSnap, LeafSnap, Plant Identifier and Pl@ntNet™. The applications were chosen based on ease of availability or whether they had been previously researched.

The researchers were allocated two of the applications and asked to download each application onto their personal telephone, via either the Apple AppStore or the Google Play Store. The smartphones used belonged to the individual researchers and were either Apple devices using iOS 16 or Android devices using version 13. There was no commercial sponsorship for this study. The apps used were either at no cost or used during a free trial period. Each researcher was given access to a secure cloud-based folder containing de-identified copies of the images from each exposure. The researchers were asked to identify the images as though part of a real poisons consultation to mimic how they would use them in a real-world scenario. They were then requested to record their most likely identification for each image, based on the results of each application, and finally the most likely single identification for each exposure. The recorded answers were compared to the identification from the botanist. A match was determined if the genus was correct, using either scientific or common names.

Outcome

The primary outcome was the proportion of correct identification by the plant identifier application for each poisoning exposure. Secondary outcomes were the proportion of correct identification for each individual image and the agreement between users of each application.

Analysis

Descriptive statistics were used. We calculated the proportion of correct identification with 95% confidence intervals. Cohan’s kappa coefficient was calculated to determine agreement. All analysis was performed in GraphPad Prism 10.1 or Mac OS X (GraphPad Software, San Diego, CA, USA; www.graph-pad.com).

Results

During the study period, QPIC received images for 25 plant exposures, and the caller was directed to the Queensland Herbarium for a positive identification. In most cases, more than one image was received, resulting in a total of 73 images for analysis. Recruitment was stopped after 12 months due to funding constraints with fewer cases recruited than expected.

Most exposures were in males (17/25, 68%), under 14 years old (22/25, 88%), and occurring at home (19/25, 76%). There were five cases with mild symptoms including gastrointestinal or skin irritation. There were no cases having moderate or severe symptoms (Table 1). There were 18 different genera of plant among the 25 exposures. Eight of the plants are considered non-toxic, and two contain toxic alkaloids such as solanine or pyrrolizidine alkaloids. Eight could cause skin or gastrointestinal irritation including plants containing oxalates (Table 1).

Table 1. Baseline characteristics of 25 plant-related exposures referred to the Queensland Poisons Information Centre.

Characteristic		Number	Expected toxin
Sex	Male	17
Age	<12 months	2
	1–4 years	14
	5–14 years	6
	>14 years	3
Intent	Intentional	4
	Unintentional	21
Location	Home	19
	Educational	5
	Other	1
Exposure	Ingested	22
	Eyes	2
	Dermal	1
Plant	Solanine spp.	3	Glycoalkaloids
	Rivina humulis	3	Gastrointestinal/skin irritants
	Ficus spp.	2	Gastrointestinal/skin irritants
	Szyigium australe	2	Non-toxic
	Dietes spp.	2	Non-toxic
	Spathodea campanulata	2	Non-toxic
	Senecio madagascariensis	1	Pyrrolizidine alkaloids
	Duranta erecta	1	Gastrointestinal/skin irritants
	Koelreuteria paniculata	1	Gastrointestinal/skin irritants
	Asparagus aethiopicus	1	Gastrointestinal/skin irritants
	Hippeastrum puniceum	1	Gastrointestinal/skin irritants
	Philodendron	1	Oxalates
	Epipremnum pinnatum	1	Oxalates
	Archontophoenix spp.	1	Non-toxic
	Sonchus oleraceus	1	Non-toxic
	Ardisia japonica	1	Non-toxic
	Liriope	1	Non-toxic
	Chamaedorea	1	Non-toxic
Poisoning Severity Score	0	20
	1	5
	2	0
	3	0
	4	0

Outcomes

Table 2 presents the accuracy of application identification. There was a wide range of variability between the best-performing apps (PictureThis 74%, Pl@ntnet™ and Plant Identifier both 72%) and the poorest-performing app (PlantSnap 38%).

Table 2. Study outcomes for accuracy and agreement in using plant identification applications for plant poisoning exposures, total numbers reflecting two reviewers per case or image.

Application	Proportion of correct identification per exposure (95% CI) n = 50	Proportion of correct identification per image (95% CI) n = 146	Agreement between application users (kappa) (95% CI)
Google Lens™	60% (45–74%)	49% (41–58%)	0.68 (0.40–0.95)
PictureThis	74% (60–85%)	64% (56–72%)	0.44 (0.12–0.77)
PlantSnap	38% (25–53%)	23% (16–30%)	0.58 (0.25–0.91)
LeafSnap	60% (45–74%)	53% (44–61%)	0.51 (0.19–0.83)
Plant Identifier	72% (58–84%)	51% (42–59%)	0.80 (0.55–1.0)
Pl@ntnet^™	72% (58–84%)	68% (60–75%)	1.0 (1.0–1.0)

CI: confidence interval.

The accuracy of identifying individual images was relatively lower compared to the overall identification given to each exposure. This was true for all the tested apps, with Pl@ntNet™ performing the best at 68% and PlantSnap having the lowest performance at only 28%.

There was also variability in how well individual images could be identified, with certain images performing very well across multiple apps (Figure 1) and others performing very poorly across all apps (Figure 2).

Figure 1. Examples of images with high identification accuracy. 1. Dietes bicolor. 2. Duranta spp. (golden dewdrop). 3. Philodendron.

Figure 2. Examples of images with poor identification accuracy. 1. Syzigium australe. 2. Liriope. 3. Epipremnum. 4. Solanum americanum.

Regarding two reviewers using each application, the interrater reliability followed a similar pattern to accuracy. Pl@ntnet™ had a remarkably high degree of agreement (kappa of 1) to the level of genus, although there were some discrepancies with individual species. Plant Identifier (0.8) and Google Lens™ (0.68) also had a high degree of agreement, with remaining applications only a moderate degree.

Discussion

While plant exposures are a common call to poisons information centres, severe effects are rare. Early identification of the plants involved can help discern which exposures are at risk of toxicity and need medical review and those who can remain in the community.

Deep neural networks are a form of artificial intelligence that can be trained using large datasets. They are a rising image identification method and are being increasingly investigated in medical [14] and other areas [7,9]. The widespread availability of smartphones, with their high-resolution cameras and processing power, has made these techniques accessible to a broader population. The accuracy of these techniques has been investigated for identifying healthy and diseased plants and their utility in identifying poisonous plant species and mushrooms.

Prior studies have looked at the accuracy of plant identification apps with a view to their ability to identify toxic varieties using images supplied for research purposes. Otter et al. found variable accuracy when comparing three apps (PictureThis 96%, Pl@ntnet™ 91% and PlantSnap 56%) when using photographs of toxic plants found in their locality explicitly taken for research purposes. Mahonski et al. found that PictureThis had an 82% accuracy (at the genus level) when using photographs taken from textbooks, and misidentifications occurred with some significantly toxic plants such as Conium maculatum. Long et al. again used photographs taken for research purposes and included toxic, potentially toxic and edible plants. They found an overall accuracy of 77.7% with considerable variability between the applications (LeafSnap 70%, PictureThis 96%, Pl@ntnet™ 86%, PlantSnap 53%) with misidentifications of toxic plants occurring. Google Lens™ was chosen to study due to its availability and large installed user base. Its accuracy in this study is similar to other studies in which it is used for plant identification, and it is interesting that in both this study and others, it performs as well or better than some dedicated plant identification applications [15].

The best accuracy in our study was 74% (PictureThis), and overall, the accuracy was lower in this study vs the prior work. This may reflect the quality of the supplied photographs, which were taken by members of the public rather than experts. The variable quality of the images can be seen by comparing images that are most frequently identified correctly (Figure 1), in which you can see good quality images of large parts of plants, including whole leaves, berries, and flowers. In contrast, those with the lowest likelihood of correct identification (Figure 2) show parts of damaged plants or fruit. The impressive ability of the botanists to identify even poor-quality images suggests that additional factors may be at play, such as experience or that additional helpful information is gained when being consulted, such as location or additional identifying features of the plant. In one example, images from two different plants were supplied for identification. The botanist correctly identified both plants, while the apps identified one or the other, and the researcher picked the one they felt most likely.

The accuracy between plant-identifying applications varied greatly. PictureThis was the most accurate at 74%, followed by Plant Identifier and Pl@ntnet™ at 72%. PlantSnap was the least reliable, with an accuracy rate of only 38%. There was also variability in agreement between users, with generally better interrater agreement between the better-performing apps. The two researchers using Pl@ntnet™ had a complete agreement, while the other apps had a good or moderate agreement. However, there was an exception, with PictureThis having the best overall accuracy for identification for each case but the worst interrater reliability. When identifying plants, the applications would provide a number of possible answers in the order of likelihood, although this was not always clear. The users were asked to choose the result they thought the most accurate, as this would more closely reflect clinical practice, rather than just picking the top answer. The variation in interrater reliability likely reflects both human factors related to individual choices and the variability in results from different applications.

The spectrum of calls was similar to data previously published by the Queensland Poisons Information Centre. While plant exposures are a common call to Poison Centres, most cases are either asymptomatic or mild, and only 1% are considered moderate or severe (PSS 2 or 3), with 15% recommended to seek medical attention [6]. In our study, there were no moderate or severe poisonings and none of the cases needed to be referred to a clinical toxicologist for further advice thus the performance of the applications would have had limited clinical bearing.

This study had multiple limitations. Recruitment was lower than expected, and ultimately, the sample was small. The lack of dedicated research capability at Queensland Poisons Information Centre likely contributed to this.

Identification apps are becoming more common on smartphone app stores. The apps used here re­present only a fraction and were chosen based on availability and previous studies. At the time of the study’s completion, Plant Identifier, one of the better-performing apps, was no longer available on the Australian app store. The datasets of the applications vary in their level of localisation. PictureThis, Pl@ntnet™, and PlantSnap include Australian plants, while LeafSnap does not. Therefore, the performance of the applications may differ when used in different environments.

Conclusion

Current plant identification apps are not sufficiently reliable in real-life poisoning exposures to be used by a Poisons Information Centre to help guide advice when dealing with plant exposure calls. Botanist identification remains the gold standard. Clinicians should be wary of their use when providing clinical advice.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data supporting this study’s findings are available from the corresponding author upon reasonable request.

Additional information

Funding

The study was possible because of PA Research Foundation 2022 Emergency Medicine Grant funding.